{"docs":[{"id":204,"title":"Recovering a Troubled Engineering Project: Warning Signs and First Steps","heading":"Recovering a Troubled Engineering Project: Warning Signs and First Steps","subheading":null,"category":"Consulting","body":{"root":{"type":"root","children":[{"type":"heading","tag":"h3","version":1,"children":[{"type":"text","text":"The first steps","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Freeze new scope and establish the real state of the project independently: build the software from source, run the tests, measure the hardware against its requirements and list every open defect by severity.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Then decide, subsystem by subsystem, whether to fix, simplify, descope or redesign. Rebuild the plan from that baseline, with the riskiest work first, and agree it openly with the sponsor. Report against the new plan frequently, so confidence is rebuilt on results.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"tags":"Consulting, Project Recovery, Risk Management","cardCta":{"label":null,"link":null},"pageCta":{"label":"Talk to AAVICO engineering","link":"/contact"},"excerpt":"Troubled projects give warning long before they fail. Recognising the signs and taking the right first steps makes the difference between a recovery and a write-off.","featuredImage":null,"type":"Blog","readTime":"3 min read","publishedAt":"2026-11-15T13:19:00.000Z","heroImage":null,"introText":"Engineering projects rarely fail suddenly. They drift: dates slip a week at a time, integration is postponed, and the same tasks are reported as nearly done for months. By the time the problem is acknowledged, the team is tired and the sponsor has lost confidence. Recovery is possible, but it starts with an honest view of where the project really stands. This article lists the warning signs and the first steps that work.","metrics":[{"id":"recovering-a-troubled-engineering-project-warnin-metric-1","value":"90% done","label":"A task reported nearly complete for weeks is a classic warning sign.","icon":null},{"id":"recovering-a-troubled-engineering-project-warnin-metric-2","value":"Freeze","label":"Pausing new scope so the team can stabilise what exists.","icon":null},{"id":"recovering-a-troubled-engineering-project-warnin-metric-3","value":"Baseline","label":"An honest record of what works, what does not and what is unknown.","icon":null},{"id":"recovering-a-troubled-engineering-project-warnin-metric-4","value":"Re-plan","label":"A schedule rebuilt from evidence, not from the original date.","icon":null}],"gallery":[],"challenge":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Warning signs are usually visible early: integration deferred to the end, a growing list of open defects, requirements changing without any assessment of impact, and demonstrations that only work on one unit.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"The natural responses often make things worse. Adding people to a late project tends to slow it at first, because newcomers need time from the existing team, and pushing harder on the old plan repeats the old result. Quietly cutting testing to save time simply moves the problems into the field.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"blockquote":{"quote":"Recovery starts with an honest baseline. You cannot re-plan from a position you have not measured.","author":"","role":"AAVICO Engineering"},"supportImage1":null,"deliverables":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Pause scope changes and gather the facts: what builds, what passes its tests and what has never been tried.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Triage defects and risks, find the true critical path and decide what can be descoped.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Publish a new plan with short, demonstrable milestones, and track progress by results rather than effort.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"results":[{"id":"recovering-a-troubled-engineering-project-warnin-result-1","title":"Clear Baseline","description":"Everyone works from the same honest view of the project."},{"id":"recovering-a-troubled-engineering-project-warnin-result-2","title":"Credible Plan","description":"Dates are based on measured progress."},{"id":"recovering-a-troubled-engineering-project-warnin-result-3","title":"Restored Confidence","description":"Short milestones show the sponsor real results."},{"id":"recovering-a-troubled-engineering-project-warnin-result-4","title":"Lessons Kept","description":"The causes of the trouble are understood and addressed."}],"lookingAhead":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"An independent reviewer can establish the baseline faster and with less friction than a team defending its own work.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Once the project is stable, hold a review of what caused the trouble so the same patterns do not recur.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"supportImage2":null,"testimonial":{"quote":null,"author":null,"role":null,"photo":null},"challengesSolved":[{"id":"recovering-a-troubled-engineering-project-warnin-challenge-1","title":"Integration left until the end","description":null,"icon":null},{"id":"recovering-a-troubled-engineering-project-warnin-challenge-2","title":"Tasks stuck at nearly done","description":null,"icon":null},{"id":"recovering-a-troubled-engineering-project-warnin-challenge-3","title":"Scope changes with no impact assessment","description":null,"icon":null},{"id":"recovering-a-troubled-engineering-project-warnin-challenge-4","title":"People added late to a late project","description":null,"icon":null}],"relatedPosts":[],"contactUs":{"show":true},"updatedAt":"2026-11-15T17:32:32.682Z","createdAt":"2026-11-15T17:32:32.682Z"},{"id":203,"title":"Engineering Project Audits: What an Independent Review Looks For","heading":"Engineering Project Audits: What an Independent Review Looks For","subheading":null,"category":"Consulting","body":{"root":{"type":"root","children":[{"type":"heading","tag":"h3","version":1,"children":[{"type":"text","text":"How a review is carried out","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Reviewers interview the team, sample the project records and trace a selection of requirements through design to the tests that verify them. Gaps in that chain are usually more revealing than any single document. Reviewers also confirm that the version being tested is the version being reported on, and that suppliers and manufacturing are ready for the next stage.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"The report grades each finding by severity, explains its consequence and recommends a proportionate fix. A good audit also records what is working well, so the team knows what to keep doing.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"tags":"Consulting, Project Audit, Quality","cardCta":{"label":null,"link":null},"pageCta":{"label":"Talk to AAVICO engineering","link":"/contact"},"excerpt":"An independent audit checks whether a project is on track using evidence rather than status reports. Here is what a reviewer examines and when an audit is most useful.","featuredImage":null,"type":"Blog","readTime":"3 min read","publishedAt":"2026-11-15T13:26:00.000Z","heroImage":null,"introText":"Project status reports are written by the people doing the work, and they tend to report effort rather than results. An independent audit looks at the evidence directly: the requirements, the design decisions, the test results and the plan. Its purpose is not to assign blame but to find gaps while they are still cheap to close. This article describes what a review covers, how it is carried out and when to commission one.","metrics":[{"id":"engineering-project-audits-what-an-independent-r-metric-1","value":"Independent","label":"Reviewers with no stake in the design or schedule being assessed.","icon":null},{"id":"engineering-project-audits-what-an-independent-r-metric-2","value":"Traceability","label":"Requirements linked to design and tests, so gaps become visible.","icon":null},{"id":"engineering-project-audits-what-an-independent-r-metric-3","value":"Evidence","label":"Findings drawn from documents, data and tests, not from status reports.","icon":null},{"id":"engineering-project-audits-what-an-independent-r-metric-4","value":"Graded","label":"Findings ranked by severity so effort goes where risk is highest.","icon":null}],"gallery":[],"challenge":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Teams reviewing their own work share its assumptions, so the same blind spots survive every internal review. Status can stay green for months while requirements remain untested and integration is postponed.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"The cost of a gap grows the later it is found. A missing requirement found at design review costs a document change; found at certification or in the field, it can cost a redesign.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"blockquote":{"quote":"An audit should tell you what is true about the project, especially when it differs from the plan.","author":"","role":"AAVICO Engineering"},"supportImage1":null,"deliverables":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Agree the scope, the stage of the project, the questions the sponsor needs answered and the records the team will make available.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Review the evidence, interview the team and trace sample requirements end to end.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Present graded findings and recommendations, then agree owners and dates for the actions.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"results":[{"id":"engineering-project-audits-what-an-independent-r-result-1","title":"Early Detection","description":"Gaps are found while they are still inexpensive to close."},{"id":"engineering-project-audits-what-an-independent-r-result-2","title":"Honest Status","description":"Sponsors see progress measured by results."},{"id":"engineering-project-audits-what-an-independent-r-result-3","title":"Focused Action","description":"Severity grading points effort at the real risks."},{"id":"engineering-project-audits-what-an-independent-r-result-4","title":"Stronger Process","description":"Recurring issues are fixed at their cause."}],"lookingAhead":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Schedule audits at decision points such as design freeze or before committing to tooling, when findings can still change the outcome. A short review at the right moment is worth more than a long one after the decision.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Share findings with the team directly and early. Audits work best when engineers see them as help rather than inspection.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"supportImage2":null,"testimonial":{"quote":null,"author":null,"role":null,"photo":null},"challengesSolved":[{"id":"engineering-project-audits-what-an-independent-r-challenge-1","title":"Requirements with no verifying test","description":null,"icon":null},{"id":"engineering-project-audits-what-an-independent-r-challenge-2","title":"Design decisions with no recorded reason","description":null,"icon":null},{"id":"engineering-project-audits-what-an-independent-r-challenge-3","title":"Builds that cannot be reproduced","description":null,"icon":null},{"id":"engineering-project-audits-what-an-independent-r-challenge-4","title":"A risk register nobody updates","description":null,"icon":null}],"relatedPosts":[],"contactUs":{"show":true},"updatedAt":"2026-11-15T17:32:32.682Z","createdAt":"2026-11-15T17:32:32.682Z"},{"id":202,"title":"Technical Due Diligence for Hardware and Software Products","heading":"Technical Due Diligence for Hardware and Software Products","subheading":null,"category":"Consulting","body":{"root":{"type":"root","children":[{"type":"heading","tag":"h3","version":1,"children":[{"type":"text","text":"What reviewers examine","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"For hardware: native schematic and layout files, the bill of materials with lifecycle status and alternate sources, test and certification reports, and manufacturing documentation. For software: a clean build from version control, the dependency list with licences, test coverage, security practice and the update mechanism.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Copyleft licences such as the GPL carry obligations when the software is distributed, so an inventory of open-source components matters. Findings are graded by severity, with an estimate of the effort needed to resolve each one.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"tags":"Consulting, Due Diligence, Risk Assessment","cardCta":{"label":null,"link":null},"pageCta":{"label":"Talk to AAVICO engineering","link":"/contact"},"excerpt":"Before investing in, acquiring or licensing a product, an independent technical review shows what actually exists behind the demonstration and what it will cost to fix.","featuredImage":null,"type":"Blog","readTime":"3 min read","publishedAt":"2026-11-15T13:33:00.000Z","heroImage":null,"introText":"A product demonstration shows what works on the day. Technical due diligence looks at what sits behind it: whether the design files exist and are complete, whether the software can be rebuilt, whether key parts are still in production and whether the code can legally be shipped. It is used by investors, acquirers and companies taking over a product from a supplier. This article explains what a review examines and the warning signs to look for.","metrics":[{"id":"technical-due-diligence-for-hardware-and-softwar-metric-1","value":"Ownership","label":"Who holds the designs, code and licences the product depends on.","icon":null},{"id":"technical-due-diligence-for-hardware-and-softwar-metric-2","value":"Licences","label":"Open-source obligations that follow the code into the product.","icon":null},{"id":"technical-due-diligence-for-hardware-and-softwar-metric-3","value":"Key people","label":"How much of the product exists only in a few engineers’ heads.","icon":null},{"id":"technical-due-diligence-for-hardware-and-softwar-metric-4","value":"Lifecycle","label":"Whether key components are active, not recommended for new designs, or end of life.","icon":null}],"gallery":[],"challenge":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"A product can work well and still be expensive to own. Common hidden problems include design files held only as PDFs, firmware that builds on one machine, components nearing end of life and certification that was assumed but never completed.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"These issues rarely show in a demonstration or a slide deck. They appear after the transaction, when the new owner tries to manufacture, modify or certify the product.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"blockquote":{"quote":"Ask to rebuild the product from its files. What cannot be rebuilt is not really owned.","author":"","role":"AAVICO Engineering"},"supportImage1":null,"deliverables":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Agree the scope and the questions the review must answer, then issue a document and access request.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Examine the evidence hands-on: rebuild the software, inspect design files and interview the engineers who know the product.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Report findings graded by severity, with the cost and time to resolve each, so they can inform the decision.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"results":[{"id":"technical-due-diligence-for-hardware-and-softwar-result-1","title":"Clear Picture","description":"The real state of the product is known before commitment."},{"id":"technical-due-diligence-for-hardware-and-softwar-result-2","title":"Priced Risk","description":"Problems carry effort estimates that feed into the decision."},{"id":"technical-due-diligence-for-hardware-and-softwar-result-3","title":"Smoother Handover","description":"Gaps in documentation are found while the original team is available."},{"id":"technical-due-diligence-for-hardware-and-softwar-result-4","title":"Legal Clarity","description":"Licence obligations are identified before shipping."}],"lookingAhead":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Legal and licence questions identified in the review should be confirmed by a qualified lawyer. A technical review identifies them; it does not give legal advice.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"If the product will move to a new team or factory, plan the knowledge transfer while the original engineers are still involved.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"supportImage2":null,"testimonial":{"quote":null,"author":null,"role":null,"photo":null},"challengesSolved":[{"id":"technical-due-diligence-for-hardware-and-softwar-challenge-1","title":"Design files that exist only as PDFs","description":null,"icon":null},{"id":"technical-due-diligence-for-hardware-and-softwar-challenge-2","title":"Firmware that builds on one machine","description":null,"icon":null},{"id":"technical-due-diligence-for-hardware-and-softwar-challenge-3","title":"Open-source obligations nobody tracked","description":null,"icon":null},{"id":"technical-due-diligence-for-hardware-and-softwar-challenge-4","title":"Certification assumed but never done","description":null,"icon":null}],"relatedPosts":[],"contactUs":{"show":true},"updatedAt":"2026-11-15T17:32:32.682Z","createdAt":"2026-11-15T17:32:32.682Z"},{"id":201,"title":"Feasibility Studies: Testing an Idea Before Committing a Budget","heading":"Feasibility Studies: Testing an Idea Before Committing a Budget","subheading":null,"category":"Consulting","body":{"root":{"type":"root","children":[{"type":"heading","tag":"h3","version":1,"children":[{"type":"text","text":"What a study covers","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Technical feasibility asks whether the core function can be achieved with available parts and methods, tested on the bench under realistic conditions. Commercial feasibility asks whether the unit cost, compliance requirements and component supply fit the business case.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"The deliverable is a short report: each question with its evidence and answer, the risks that remain, a recommended architecture and a rough estimate of development effort. A no-go, or a redirect, is a valid and valuable result.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"tags":"Consulting, Feasibility, Risk Reduction","cardCta":{"label":null,"link":null},"pageCta":{"label":"Talk to AAVICO engineering","link":"/contact"},"excerpt":"A feasibility study answers the riskiest questions about a product with evidence before full development is funded. Here is what one covers and how to keep it focused.","featuredImage":null,"type":"Blog","readTime":"3 min read","publishedAt":"2026-11-15T13:40:00.000Z","heroImage":null,"introText":"Most products rest on a few assumptions that, if wrong, sink the project: that a sensor can detect the signal, that a model reaches useful accuracy, that the battery lasts long enough or that the unit cost fits the price. A feasibility study tests those assumptions with experiments and estimates, within a fixed time and budget, and ends in a decision. This article explains what a good study covers and what it should deliver.","metrics":[{"id":"feasibility-studies-testing-an-idea-before-commi-metric-1","value":"Unknowns","label":"Questions nobody on the team can yet answer with evidence.","icon":null},{"id":"feasibility-studies-testing-an-idea-before-commi-metric-2","value":"Bench test","label":"A focused experiment that answers one question, not a miniature product.","icon":null},{"id":"feasibility-studies-testing-an-idea-before-commi-metric-3","value":"BOM est.","label":"An early bill-of-materials estimate that tests whether the unit cost works.","icon":null},{"id":"feasibility-studies-testing-an-idea-before-commi-metric-4","value":"Go / no-go","label":"Criteria agreed in advance, so the result leads to a decision.","icon":null}],"gallery":[],"challenge":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Without a study, the riskiest assumption is often tested last, after enclosures, firmware and app are well advanced. When it fails, most of that work is lost, along with the time and the confidence of whoever funded it.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Studies go wrong in the opposite way too. Without defined questions and an end date they drift into open-ended research, or into building a prototype product that answers nothing clearly.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"blockquote":{"quote":"A feasibility study is finished when it can support a decision, not when the prototype looks good.","author":"","role":"AAVICO Engineering"},"supportImage1":null,"deliverables":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"List the assumptions the product depends on and rank them by how likely they are to be wrong and how costly that would be.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Design the smallest experiment or estimate that answers each high-risk question, and agree pass criteria before starting.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Run the work within a fixed time box, then review the evidence and decide to proceed, change direction or stop.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"results":[{"id":"feasibility-studies-testing-an-idea-before-commi-result-1","title":"Informed Decision","description":"Funding follows evidence rather than optimism."},{"id":"feasibility-studies-testing-an-idea-before-commi-result-2","title":"Retired Risk","description":"The biggest unknowns are answered first, when changes are cheap."},{"id":"feasibility-studies-testing-an-idea-before-commi-result-3","title":"Realistic Estimate","description":"Development effort and unit cost are grounded in measurement."},{"id":"feasibility-studies-testing-an-idea-before-commi-result-4","title":"Better Architecture","description":"Experiments shape the design before it is fixed."}],"lookingAhead":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"If the study proceeds to development, carry its experiments forward as early test cases. They often become the first verification tests.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Keep the study code and hardware clearly separate from the product design. Feasibility work is built for speed, not for production.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"supportImage2":null,"testimonial":{"quote":null,"author":null,"role":null,"photo":null},"challengesSolved":[{"id":"feasibility-studies-testing-an-idea-before-commi-challenge-1","title":"The riskiest assumption tested last","description":null,"icon":null},{"id":"feasibility-studies-testing-an-idea-before-commi-challenge-2","title":"A study with no end date","description":null,"icon":null},{"id":"feasibility-studies-testing-an-idea-before-commi-challenge-3","title":"Results with no agreed pass criteria","description":null,"icon":null},{"id":"feasibility-studies-testing-an-idea-before-commi-challenge-4","title":"Prototype code shipped as product code","description":null,"icon":null}],"relatedPosts":[],"contactUs":{"show":true},"updatedAt":"2026-11-15T17:32:32.682Z","createdAt":"2026-11-15T17:32:32.682Z"},{"id":200,"title":"ROS 2 for Robotics Products: Architecture and Practical Use","heading":"ROS 2 for Robotics Products: Architecture and Practical Use","subheading":null,"category":"Robotics","body":{"root":{"type":"root","children":[{"type":"heading","tag":"h3","version":1,"children":[{"type":"text","text":"Topics, services and actions","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Topics carry continuous streams such as sensor data and commands, from any number of publishers to any number of subscribers. Services handle short request-and-response exchanges. Actions handle long-running goals with feedback and cancellation, and are built from topics and services.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"The middleware is pluggable through the ROS middleware interface, with several DDS implementations and, in recent distributions, Zenoh available. Behaviour and tuning differ between implementations, so choose one, test with it and keep it fixed for the product. Composing several nodes into one process can also cut copying and latency for high-rate data.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"tags":"Robotics, ROS 2, Software Architecture","cardCta":{"label":null,"link":null},"pageCta":{"label":"Talk to AAVICO engineering","link":"/contact"},"excerpt":"ROS 2 gives robots a common structure for drivers, planning and control. Using it in a product means understanding its communication patterns, middleware and quality of service.","featuredImage":null,"type":"Blog","readTime":"3 min read","publishedAt":"2026-11-15T13:47:00.000Z","heroImage":null,"introText":"ROS 2 is an open-source framework that structures robot software as many small programs, called nodes, that exchange data through a middleware layer. It provides drivers, navigation, manipulation and tooling that would take years to build alone. Moving from a research prototype to a product, though, needs attention to architecture, communication settings and security. This article explains the core concepts and the practical points that matter for shipping systems.","metrics":[{"id":"ros-2-for-robotics-products-architecture-and-pra-metric-1","value":"Nodes","label":"Units of computation that each do one job, such as a driver, planner or controller.","icon":null},{"id":"ros-2-for-robotics-products-architecture-and-pra-metric-2","value":"DDS","label":"The publish-subscribe middleware standard ROS 2 uses for discovery and transport by default.","icon":null},{"id":"ros-2-for-robotics-products-architecture-and-pra-metric-3","value":"QoS","label":"Quality of service settings for reliability, durability and history depth.","icon":null},{"id":"ros-2-for-robotics-products-architecture-and-pra-metric-4","value":"Lifecycle","label":"Managed nodes with defined states for controlled start-up and shutdown.","icon":null}],"gallery":[],"challenge":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Prototype systems tend to grow into dozens of nodes started by sprawling launch files, with settings scattered across them. Start-up order becomes fragile and nobody is sure which configuration is running on a given robot.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Communication problems can be silent. A subscriber requesting reliable delivery will not connect to a best-effort publisher, and robots sharing a network discover each other unless their domains are separated.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"blockquote":{"quote":"Treat topics and their QoS settings as interfaces. Design them, document them and version them.","author":"","role":"AAVICO Engineering"},"supportImage1":null,"deliverables":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Map the system as nodes and interfaces, choosing topics, services or actions for each exchange and setting QoS deliberately.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Use lifecycle nodes and centralised parameter files so start-up, configuration and shutdown are predictable.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Pin a long-term support distribution and your dependencies, and enable security if the robot shares a network.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"results":[{"id":"ros-2-for-robotics-products-architecture-and-pra-result-1","title":"Clear Structure","description":"Each node has one job and documented interfaces."},{"id":"ros-2-for-robotics-products-architecture-and-pra-result-2","title":"Reliable Start-Up","description":"Managed lifecycles bring the system up in a known order."},{"id":"ros-2-for-robotics-products-architecture-and-pra-result-3","title":"Reuse","description":"Standard drivers and packages shorten development."},{"id":"ros-2-for-robotics-products-architecture-and-pra-result-4","title":"Maintainability","description":"Pinned versions keep every robot running known software."}],"lookingAhead":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Use ROS_DOMAIN_ID or network separation to keep robots and test benches from discovering each other unintentionally.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Keep hard real-time control in dedicated firmware or a real-time controller, with ROS 2 commanding it at a higher level. This keeps tight loops independent of the operating system scheduler and the network.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"supportImage2":null,"testimonial":{"quote":null,"author":null,"role":null,"photo":null},"challengesSolved":[{"id":"ros-2-for-robotics-products-architecture-and-pra-challenge-1","title":"Silent QoS mismatches between nodes","description":null,"icon":null},{"id":"ros-2-for-robotics-products-architecture-and-pra-challenge-2","title":"Robots discovering each other on a shared network","description":null,"icon":null},{"id":"ros-2-for-robotics-products-architecture-and-pra-challenge-3","title":"Unpinned dependencies changing between builds","description":null,"icon":null},{"id":"ros-2-for-robotics-products-architecture-and-pra-challenge-4","title":"Security left disabled in the field","description":null,"icon":null}],"relatedPosts":[],"contactUs":{"show":true},"updatedAt":"2026-11-15T17:32:32.682Z","createdAt":"2026-11-15T17:32:32.682Z"},{"id":199,"title":"Robot Motion Control: Motors, Encoders and Control Loops","heading":"Robot Motion Control: Motors, Encoders and Control Loops","subheading":null,"category":"Robotics","body":{"root":{"type":"root","children":[{"type":"heading","tag":"h3","version":1,"children":[{"type":"text","text":"Cascaded loops","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"The innermost current loop sets motor torque and usually runs in the motor drive. A velocity loop around it commands current, and a position loop around that commands velocity. Each inner loop should be considerably faster than the loop outside it so the outer loop sees it as nearly ideal.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Add feedforward of velocity and acceleration from the trajectory to reduce tracking lag, limit the integrator when the output saturates, and filter or take the derivative from the measurement rather than the error.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"tags":"Robotics, Motion Control, PID","cardCta":{"label":null,"link":null},"pageCta":{"label":"Talk to AAVICO engineering","link":"/contact"},"excerpt":"Smooth, accurate robot motion comes from the right motor, good feedback and well-tuned control loops. Here is how the pieces fit together and where projects go wrong.","featuredImage":null,"type":"Blog","readTime":"3 min read","publishedAt":"2026-11-15T13:54:00.000Z","heroImage":null,"introText":"Every robot joint or wheel is a small control system: a motor produces torque, a sensor measures what happened and a controller corrects the difference. The usual structure is a set of nested loops controlling current, velocity and position, each typically using a form of PID control. This article explains the main motor and encoder choices and how the loops are arranged and tuned.","metrics":[{"id":"robot-motion-control-motors-encoders-and-control-metric-1","value":"PID","label":"Proportional, integral and derivative terms acting on the error between target and measurement.","icon":null},{"id":"robot-motion-control-motors-encoders-and-control-metric-2","value":"Cascade","label":"Current, velocity and position loops nested, with the inner loop fastest.","icon":null},{"id":"robot-motion-control-motors-encoders-and-control-metric-3","value":"Quadrature","label":"Two encoder channels a quarter cycle apart give both position and direction.","icon":null},{"id":"robot-motion-control-motors-encoders-and-control-metric-4","value":"4x","label":"Counting every edge on both quadrature channels gives four counts per encoder line.","icon":null}],"gallery":[],"challenge":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Control problems are often mechanical problems in disguise. Backlash in a gearbox, friction that changes with temperature and flexible couplings all limit how tightly a loop can be tuned, and no gain setting removes them. Resonances in the structure set a further ceiling on loop bandwidth.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Others come from the controller. A loop that runs too slowly, noisy derivative terms and integrators that wind up while the motor is saturated all cause overshoot, oscillation or sluggish response.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"blockquote":{"quote":"Tune from the inside out. An outer loop can never be better than the loop it commands.","author":"","role":"AAVICO Engineering"},"supportImage1":null,"deliverables":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Size the motor and gearing from the required torque, speed and inertia, with margin for acceleration.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Choose feedback resolution from the positioning accuracy, and decide whether absolute position is needed at power-up.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Tune the current, velocity and position loops in that order, then add feedforward and test across the full load range.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"results":[{"id":"robot-motion-control-motors-encoders-and-control-result-1","title":"Accuracy","description":"The axis reaches its target position repeatably."},{"id":"robot-motion-control-motors-encoders-and-control-result-2","title":"Smooth Motion","description":"Tuned loops and feedforward reduce overshoot and vibration."},{"id":"robot-motion-control-motors-encoders-and-control-result-3","title":"Robustness","description":"Behaviour stays stable as load and temperature vary."},{"id":"robot-motion-control-motors-encoders-and-control-result-4","title":"Safe Behaviour","description":"Saturation, limits and faults are handled predictably."}],"lookingAhead":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Incremental encoders need a homing move after power-up; absolute encoders report position immediately. Choose according to whether the robot can safely move to find home.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Plan safety functions such as torque limits and safe stopping early, since they shape the drive and the control architecture.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"supportImage2":null,"testimonial":{"quote":null,"author":null,"role":null,"photo":null},"challengesSolved":[{"id":"robot-motion-control-motors-encoders-and-control-challenge-1","title":"Gains tuned around mechanical backlash","description":null,"icon":null},{"id":"robot-motion-control-motors-encoders-and-control-challenge-2","title":"Integrator windup during saturation","description":null,"icon":null},{"id":"robot-motion-control-motors-encoders-and-control-challenge-3","title":"A position loop faster than its velocity loop","description":null,"icon":null},{"id":"robot-motion-control-motors-encoders-and-control-challenge-4","title":"Homing moves that are unsafe","description":null,"icon":null}],"relatedPosts":[],"contactUs":{"show":true},"updatedAt":"2026-11-15T17:32:32.682Z","createdAt":"2026-11-15T17:32:32.682Z"},{"id":198,"title":"Object Detection Metrics: Precision, Recall and mAP Explained","heading":"Object Detection Metrics: Precision, Recall and mAP Explained","subheading":null,"category":"Computer Vision","body":{"root":{"type":"root","children":[{"type":"heading","tag":"h3","version":1,"children":[{"type":"text","text":"From boxes to mAP","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Intersection over union is the overlap area of a predicted and a true box divided by the area of their union. Detections are sorted by confidence and matched to unmatched true boxes above the IoU threshold; duplicates and unmatched detections are false positives, and unfound objects are false negatives.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Sweeping the confidence threshold traces a precision-recall curve. Average precision is the area under that curve for one class, and mAP is the mean across classes, and in COCO also across IoU thresholds.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"tags":"Computer Vision, Object Detection, Metrics","cardCta":{"label":null,"link":null},"pageCta":{"label":"Talk to AAVICO engineering","link":"/contact"},"excerpt":"Detection results are reported as precision, recall and mAP, and the numbers are easy to misread. These are the definitions and how to use them to judge a model.","featuredImage":null,"type":"Blog","readTime":"3 min read","publishedAt":"2026-11-15T14:01:00.000Z","heroImage":null,"introText":"An object detector outputs boxes, classes and confidence scores, and judging it means deciding which boxes count as correct. Precision, recall and mean average precision are the standard measures, and each answers a different question. Understanding how they are calculated makes it possible to compare models fairly and to choose a confidence threshold that suits the product rather than the leaderboard.","metrics":[{"id":"object-detection-metrics-precision-recall-and-ma-metric-1","value":"TP/(TP+FP)","label":"Precision: the share of detections that are correct.","icon":null},{"id":"object-detection-metrics-precision-recall-and-ma-metric-2","value":"TP/(TP+FN)","label":"Recall: the share of real objects that were found.","icon":null},{"id":"object-detection-metrics-precision-recall-and-ma-metric-3","value":"IoU 0.5","label":"A common overlap threshold for counting a predicted box as a match.","icon":null},{"id":"object-detection-metrics-precision-recall-and-ma-metric-4","value":"0.5:0.95","label":"The COCO convention, averaging over IoU thresholds from 0.5 to 0.95.","icon":null}],"gallery":[],"challenge":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"A single mAP figure hides the trade-off the product actually faces. A detector can score well overall while missing a rare but critical class, or while producing more false alarms than an operator will tolerate.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Figures are also computed differently. The PASCAL VOC convention uses one IoU threshold of 0.5, while COCO averages over ten thresholds, so mAP values calculated under different protocols cannot be compared. Results on a public benchmark also say little about performance on your own images.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"blockquote":{"quote":"Report precision and recall at the threshold you will ship, per class. mAP alone does not tell you that.","author":"","role":"AAVICO Engineering"},"supportImage1":null,"deliverables":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Agree the evaluation protocol, IoU threshold and test set before comparing any models.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Decide which error costs more in the product, a missed object or a false alarm, and per class if needed.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Choose the operating confidence threshold from the precision-recall curve and report both values at that point.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"results":[{"id":"object-detection-metrics-precision-recall-and-ma-result-1","title":"Fair Comparisons","description":"Models are scored under the same protocol on the same data."},{"id":"object-detection-metrics-precision-recall-and-ma-result-2","title":"Product Fit","description":"The threshold reflects the real cost of each error type."},{"id":"object-detection-metrics-precision-recall-and-ma-result-3","title":"Visible Weaknesses","description":"Per-class results expose classes the average conceals."},{"id":"object-detection-metrics-precision-recall-and-ma-result-4","title":"Clear Acceptance","description":"Pass criteria are numbers everyone understands."}],"lookingAhead":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"For counting or tracking applications, add task-level checks such as count error per image, since a good mAP does not guarantee correct counts.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Inspect false positives and false negatives by eye. Patterns in the errors, such as small, distant or partly hidden objects, usually point to missing training data.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"supportImage2":null,"testimonial":{"quote":null,"author":null,"role":null,"photo":null},"challengesSolved":[{"id":"object-detection-metrics-precision-recall-and-ma-challenge-1","title":"mAP values from different protocols compared","description":null,"icon":null},{"id":"object-detection-metrics-precision-recall-and-ma-challenge-2","title":"A rare class hidden by the average","description":null,"icon":null},{"id":"object-detection-metrics-precision-recall-and-ma-challenge-3","title":"A threshold chosen without considering cost","description":null,"icon":null},{"id":"object-detection-metrics-precision-recall-and-ma-challenge-4","title":"Recall reported without precision","description":null,"icon":null}],"relatedPosts":[],"contactUs":{"show":true},"updatedAt":"2026-11-15T17:32:32.682Z","createdAt":"2026-11-15T17:32:32.682Z"},{"id":197,"title":"Choosing Cameras and Lenses for Computer Vision","heading":"Choosing Cameras and Lenses for Computer Vision","subheading":null,"category":"Computer Vision","body":{"root":{"type":"root","children":[{"type":"heading","tag":"h3","version":1,"children":[{"type":"text","text":"Sizing the system","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Divide the field of view by the smallest feature and multiply by the pixels you want across it; a few pixels per feature is a common starting point, and measurement needs more. Motion blur in pixels is the object speed times the exposure time, divided by the object size each pixel covers.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"For focal length, a useful approximation is sensor width times working distance divided by field of view. Check the lens is rated for the sensor size and pixel pitch, and remember that a wider aperture reduces depth of field.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"tags":"Computer Vision, Cameras, Optics","cardCta":{"label":null,"link":null},"pageCta":{"label":"Talk to AAVICO engineering","link":"/contact"},"excerpt":"Image quality sets the ceiling for any vision algorithm. Resolution, shutter type, lens and lighting should be chosen from the task, not from a megapixel figure.","featuredImage":null,"type":"Blog","readTime":"3 min read","publishedAt":"2026-11-15T14:08:00.000Z","heroImage":null,"introText":"No algorithm can recover detail the camera never captured. A vision system starts with the question of what must be seen, how small it is, how fast it moves and under what light. Those answers decide the sensor resolution, the shutter type, the lens and the illumination. This article works through each choice and gives the simple calculations that size a camera for a task.","metrics":[{"id":"choosing-cameras-and-lenses-for-computer-vision-metric-1","value":"Global","label":"A global shutter exposes every pixel at once, so moving objects are not skewed.","icon":null},{"id":"choosing-cameras-and-lenses-for-computer-vision-metric-2","value":"Rolling","label":"A rolling shutter exposes rows in sequence; common and economical, but distorts fast motion.","icon":null},{"id":"choosing-cameras-and-lenses-for-computer-vision-metric-3","value":"px per mm","label":"Resolution on the object, which decides the smallest feature that can be seen.","icon":null},{"id":"choosing-cameras-and-lenses-for-computer-vision-metric-4","value":"Lighting","label":"Controlled illumination is often the cheapest way to better images.","icon":null}],"gallery":[],"challenge":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Cameras are often chosen by headline resolution. A high pixel count does not help if the lens cannot resolve it, if motion blurs the image during exposure, or if reflections and shadows change with the time of day.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Motion is a frequent cause of trouble. A rolling shutter reads rows at slightly different times, so objects moving across the frame appear slanted or wobbled, which corrupts measurements.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"blockquote":{"quote":"Size the camera from the smallest feature and the fastest motion. The megapixel count follows.","author":"","role":"AAVICO Engineering"},"supportImage1":null,"deliverables":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Define the field of view, the smallest feature, the object speed and the working distance.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Calculate the resolution and maximum exposure time, then choose the shutter type and a lens that covers the sensor.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Test candidate cameras and lighting on real parts in the real environment before freezing the mechanical design.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"results":[{"id":"choosing-cameras-and-lenses-for-computer-vision-result-1","title":"Usable Images","description":"The detail the algorithm needs is present in every frame."},{"id":"choosing-cameras-and-lenses-for-computer-vision-result-2","title":"Sharp Motion","description":"Exposure and shutter type are matched to object speed."},{"id":"choosing-cameras-and-lenses-for-computer-vision-result-3","title":"Stable Results","description":"Controlled lighting reduces sensitivity to the surroundings."},{"id":"choosing-cameras-and-lenses-for-computer-vision-result-4","title":"Right Cost","description":"Resolution and optics are sized to the task, not over-specified."}],"lookingAhead":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Consider lighting geometry as carefully as the camera: backlighting for silhouettes, diffuse light for shiny parts and low-angle light for surface texture.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Plan camera calibration if the system measures distances or positions, and protect the mount from vibration and knocks.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"supportImage2":null,"testimonial":{"quote":null,"author":null,"role":null,"photo":null},"challengesSolved":[{"id":"choosing-cameras-and-lenses-for-computer-vision-challenge-1","title":"Motion blur from a long exposure","description":null,"icon":null},{"id":"choosing-cameras-and-lenses-for-computer-vision-challenge-2","title":"Skewed images from a rolling shutter","description":null,"icon":null},{"id":"choosing-cameras-and-lenses-for-computer-vision-challenge-3","title":"A lens that cannot resolve the sensor","description":null,"icon":null},{"id":"choosing-cameras-and-lenses-for-computer-vision-challenge-4","title":"Results that change with daylight","description":null,"icon":null}],"relatedPosts":[],"contactUs":{"show":true},"updatedAt":"2026-11-15T17:32:32.682Z","createdAt":"2026-11-15T17:32:32.682Z"},{"id":196,"title":"Shrinking Models for Embedded Devices: Quantisation and Pruning","heading":"Shrinking Models for Embedded Devices: Quantisation and Pruning","subheading":null,"category":"AI & Machine Learning","body":{"root":{"type":"root","children":[{"type":"heading","tag":"h3","version":1,"children":[{"type":"text","text":"How int8 quantisation works","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Each tensor is mapped to integers with a scale and a zero point, so that real value equals scale times the integer minus the zero point. Weight ranges are known after training; activation ranges are estimated by running a representative calibration set, so poor calibration data gives poor results.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Per-channel scales for weights usually preserve accuracy better than one scale per tensor. If post-training quantisation loses too much, quantisation-aware training usually recovers most of the gap. Sensitive layers, often the first and last, can also be left at higher precision where the runtime allows mixed precision.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"tags":"AI & Machine Learning, Quantisation, Embedded AI","cardCta":{"label":null,"link":null},"pageCta":{"label":"Talk to AAVICO engineering","link":"/contact"},"excerpt":"Converting weights from 32-bit floats to 8-bit integers cuts model memory by a factor of four. Done carefully, accuracy changes little. Here is how it works.","featuredImage":null,"type":"Blog","readTime":"3 min read","publishedAt":"2026-11-15T14:15:00.000Z","heroImage":null,"introText":"Models are usually trained in 32-bit floating point, and that precision is rarely needed to run them. Quantisation stores weights and activations as small integers, and pruning removes weights that contribute little. Together they let a model fit into the memory of an embedded processor and run on integer hardware such as many NPUs. This article explains both techniques, what they cost in accuracy and when each one actually speeds things up.","metrics":[{"id":"shrinking-models-for-embedded-devices-quantisati-metric-1","value":"4x","label":"Memory saving when float32 weights are stored as int8.","icon":null},{"id":"shrinking-models-for-embedded-devices-quantisati-metric-2","value":"PTQ","label":"Post-training quantisation, which converts a trained model using a small calibration set.","icon":null},{"id":"shrinking-models-for-embedded-devices-quantisati-metric-3","value":"QAT","label":"Quantisation-aware training, which simulates int8 during training to recover accuracy.","icon":null},{"id":"shrinking-models-for-embedded-devices-quantisati-metric-4","value":"Sparsity","label":"The fraction of weights set to zero by pruning.","icon":null}],"gallery":[],"challenge":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"A model that fits comfortably on a workstation may need several times the memory an embedded part has. Many accelerators also run only integer operations, so a floating-point model cannot use them at all.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Shrinking a model is not free. Some networks lose noticeable accuracy when quantised, and pruning individual weights saves no time unless the runtime and hardware can exploit the zeros.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"blockquote":{"quote":"Measure accuracy after every compression step, on the same test set, on the target runtime.","author":"","role":"AAVICO Engineering"},"supportImage1":null,"deliverables":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Establish a baseline accuracy and latency for the float model on the agreed test set.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Apply post-training int8 quantisation with representative calibration data, then measure again on the target runtime.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"If accuracy falls short, try quantisation-aware training, structured pruning or a smaller architecture, checking each change against the baseline.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"results":[{"id":"shrinking-models-for-embedded-devices-quantisati-result-1","title":"Smaller Footprint","description":"Weights take a quarter of the memory in int8 compared with float32."},{"id":"shrinking-models-for-embedded-devices-quantisati-result-2","title":"Faster Inference","description":"Integer arithmetic suits NPUs and microcontroller instruction sets."},{"id":"shrinking-models-for-embedded-devices-quantisati-result-3","title":"Lower Power","description":"Less data moved from memory means less energy per inference."},{"id":"shrinking-models-for-embedded-devices-quantisati-result-4","title":"Known Trade-Off","description":"Accuracy loss is measured rather than assumed."}],"lookingAhead":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Knowledge distillation, training a small model to imitate a larger one, is another route to a compact model and combines well with quantisation.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Lower precisions such as int4 are increasingly supported, but support varies by runtime and hardware. Check before planning around them, and expect the accuracy cost to be larger than at int8.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"supportImage2":null,"testimonial":{"quote":null,"author":null,"role":null,"photo":null},"challengesSolved":[{"id":"shrinking-models-for-embedded-devices-quantisati-challenge-1","title":"A model too large for on-chip memory","description":null,"icon":null},{"id":"shrinking-models-for-embedded-devices-quantisati-challenge-2","title":"Calibration data that misses real conditions","description":null,"icon":null},{"id":"shrinking-models-for-embedded-devices-quantisati-challenge-3","title":"Pruning that saves no inference time","description":null,"icon":null},{"id":"shrinking-models-for-embedded-devices-quantisati-challenge-4","title":"Accuracy checked on the desktop only","description":null,"icon":null}],"relatedPosts":[],"contactUs":{"show":true},"updatedAt":"2026-11-15T17:32:32.682Z","createdAt":"2026-11-15T17:32:32.682Z"},{"id":195,"title":"Building a Training Dataset: Collection, Labelling and Bias","heading":"Building a Training Dataset: Collection, Labelling and Bias","subheading":null,"category":"AI & Machine Learning","body":{"root":{"type":"root","children":[{"type":"heading","tag":"h3","version":1,"children":[{"type":"text","text":"Where bias comes from","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Sampling bias arises when the collected data does not reflect the conditions of use. Class imbalance means rare but important cases have too few examples to learn from. Labelling bias appears when annotators interpret ambiguous cases differently or follow their own assumptions.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Counter each one deliberately: plan collection as a matrix of conditions, record metadata for every sample, label a shared subset twice to measure agreement, and review the disagreements to tighten the guidelines.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"tags":"AI & Machine Learning, Datasets, Labelling","cardCta":{"label":null,"link":null},"pageCta":{"label":"Talk to AAVICO engineering","link":"/contact"},"excerpt":"A model learns whatever its data teaches, including the shortcuts and gaps. Planning collection and labelling carefully is the most reliable way to improve results.","featuredImage":null,"type":"Blog","readTime":"3 min read","publishedAt":"2026-11-15T14:22:00.000Z","heroImage":null,"introText":"For most applied machine learning projects the dataset matters more than the choice of network. A model learns the patterns present in its examples, so missing conditions, inconsistent labels or accidental correlations all end up in its behaviour. Building a good dataset is a planned engineering activity with its own specification and quality checks. This article covers how to collect data, how to label it consistently and where bias comes from.","metrics":[{"id":"building-a-training-dataset-collection-labelling-metric-1","value":"Coverage","label":"Examples of every class and condition the product will meet in use.","icon":null},{"id":"building-a-training-dataset-collection-labelling-metric-2","value":"Guidelines","label":"Written labelling rules that make two annotators draw the same box.","icon":null},{"id":"building-a-training-dataset-collection-labelling-metric-3","value":"Agreement","label":"Measured consistency between annotators, which tests the rules themselves.","icon":null},{"id":"building-a-training-dataset-collection-labelling-metric-4","value":"Held out","label":"A test set collected and frozen before any tuning begins.","icon":null}],"gallery":[],"challenge":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Data is usually collected wherever it is easiest: one site, one camera, daytime, a few operators. A model trained on it can latch onto the background, the camera or the lighting instead of the object, and fail as soon as any of those change.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Labels carry errors too. Without clear rules, annotators disagree on where a box ends, whether a partly hidden object counts, or which class an unusual item belongs to, and the model learns that inconsistency.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"blockquote":{"quote":"Write the labelling guidelines before the first label, and expect to revise them after the first hundred.","author":"","role":"AAVICO Engineering"},"supportImage1":null,"deliverables":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"List the classes, conditions and edge cases the product must handle, and turn them into a collection plan with target counts.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Write labelling guidelines with examples of difficult cases, run a pilot batch and measure agreement between annotators.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Freeze a test set from separate sessions or sites, then label the remaining data with regular quality review.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"results":[{"id":"building-a-training-dataset-collection-labelling-result-1","title":"Better Models","description":"Consistent, representative data improves results more reliably than tuning."},{"id":"building-a-training-dataset-collection-labelling-result-2","title":"Known Gaps","description":"Metadata shows exactly which conditions are under-represented."},{"id":"building-a-training-dataset-collection-labelling-result-3","title":"Fairer Behaviour","description":"Performance is checked across groups and conditions, not just on average."},{"id":"building-a-training-dataset-collection-labelling-result-4","title":"Reusable Asset","description":"A documented dataset supports every later model version."}],"lookingAhead":{"root":{"type":"root","children":[{"type":"paragraph","version":1,"children":[{"type":"text","text":"Where images or recordings include people, check consent and data protection requirements for each jurisdiction before collection starts.","version":1}]},{"type":"paragraph","version":1,"children":[{"type":"text","text":"Consider synthetic or augmented data to fill specific gaps, but always test on real data. A model can learn the artefacts of a simulator as easily as the objects in it.","version":1}]}],"direction":"ltr","format":"","indent":0,"version":1}},"supportImage2":null,"testimonial":{"quote":null,"author":null,"role":null,"photo":null},"challengesSolved":[{"id":"building-a-training-dataset-collection-labelling-challenge-1","title":"A model that learned the background","description":null,"icon":null},{"id":"building-a-training-dataset-collection-labelling-challenge-2","title":"Annotators following different rules","description":null,"icon":null},{"id":"building-a-training-dataset-collection-labelling-challenge-3","title":"Rare cases with too few examples","description":null,"icon":null},{"id":"building-a-training-dataset-collection-labelling-challenge-4","title":"Test data drawn from the training sessions","description":null,"icon":null}],"relatedPosts":[],"contactUs":{"show":true},"updatedAt":"2026-11-15T17:32:32.682Z","createdAt":"2026-11-15T17:32:32.682Z"}],"hasNextPage":true,"hasPrevPage":false,"limit":10,"nextPage":2,"page":1,"pagingCounter":1,"prevPage":null,"totalDocs":108,"totalPages":11}