Kindly Robotics Blog
Updates on robotics for good, our mission, and what we're building — from disaster response and community cleanup to the future of physical AI.
We Timed the Loop on the Wrong Machine
A page on our own site answered in 2.5x the time of its sibling. We found the suspicious loop, timed it at 97 ms, decided 97 ms could not explain a 700 ms gap, and wrote that conclusion into a code comment. Then we shipped the fix anyway for unrelated reasons, re-measured, and the gap was gone. Here is the measurement that proved us wrong, why the ratio mattered more than the stopwatch, and the one caveat we had already written down and then ignored.
The IK Solver Returned Four Confident Answers, All 10 cm Wrong
A day spent reading our own kinematics code turned up a class of defect that is worse than a crash: a solver that answers confidently and wrongly. An inverse-kinematics branch labelled 'fall back to closest solution' returned four finite, correctly-oriented joint configurations that each put the tool 0.099 m from the pose that was asked for. Three more sites had the same shape. Here is what the numbers were, why the tests could not see it, and the honest part: most of the code we suspected turned out to be correct.
An Open Episode Interchange Format for Surgical Physical AI
Surgical AI has excellent per-task datasets but no shared episode interchange format — no equivalent of what Open X-Embodiment/RLDS, LeRobot, and DROID became for robot learning. We propose one: a four-layer episode schema that serializes to RLDS/LeRobot, ships with an installable validator, and is demonstrated on real public labels (CholecT50/Cholec80). It's a v0.1 proposal inviting critique — not an adopted standard.
Sim Isn't Enough: What the NVIDIA × NHS × Apian Story Tells Us About Clinical Physical AI
Apian's photorealistic NHS digital-twin partnership with NVIDIA — built on Project Rheo / Isaac for Healthcare — validates that hospital physical AI is now a real-money category. It also surfaces the layer simulation can't generate: credentialed, real-world clinical demonstration data. A clear-eyed analysis written pre-IRB, pre-OR-capture, pre-customer.
Why Egocentric Warehouse Data Won't Train a Surgical Robot
Mecka's EgoVerse, Build AI's Egocentric-1M, and Lightwheel's EgoSuite are racing on hour-count. For surgical robotics specifically, the implicit transfer assumption breaks across five concrete mechanisms — manipulation kinematics, viewpoint stability, tissue physics, failure-cost-shaped action distributions, and task hierarchy. A clear-eyed technical analysis of the domain gap, written pre-IRB, pre-OR-capture, pre-customer.
What HIPAA-Compliant Robot Training Data Actually Looks Like
A public design document for an IRB-cleared, HIPAA-aligned surgical data capture protocol — what it requires at the consent, pipeline, and credentialing layers, and why the funded egocentric data vendors can't ship it. Written pre-IRB, pre-contract, pre-OR-capture.
Mecka AI Alternatives: A Map of the Physical AI Data Engine Category
A neutral, vendor-by-vendor map of the physical-AI data-engine category in 2026 — egocentric collection, teleop-as-a-service, and platform/annotation plays — with verified funding, named customers, and the questions buyers should ask before signing.
No Lock-In: A Round Trip Between LeRobotDataset v3 and RLDS with FoodforThought
A practical guide to moving robot-demonstration datasets between formats: importing Hugging Face LeRobotDataset v3 and Open X-Embodiment (RLDS) into FoodforThought, and exporting labeled recordings back out to LeRobotDataset v3 and RLDS — the open standards the robot-learning ecosystem actually trains on.
A Practical Guide to Action-Segmentation Labeling for Robot-Demonstration Video
How to label robot-demonstration video accurately: frame-exact in/out points, overlap and gap rules, temporal-IoU consensus QC, and exporting clean segments to LeRobotDataset v3 and RLDS. Includes a 2-minute hands-on try-flow.