Skip to content
All work
Physical AI Safety Observability: cover image

Safety & observability

Active

Physical AI Safety Observability

Runtime safety layer on Jetson AGX Thor that measures pipeline performance and model inference, so operators can see when the system is outside safe limits.

Measured results

  • Pipeline overhead28.6 frames/sMock model, AGX Thor, 2026-09-09
  • Cosmos-Reason2-2Bp50 4.35 s per frameAGX Thor, 2026-09-09
  • False protective-gear alarms11 to 0Cosmos-Reason2-2B, required gear fields, 60 labeled simulation pictures, AGX Thor, 2026-10-17
  • People found, Cosmos 3 Edge44 of 45Against 29 for Cosmos-Reason2-2B, same pictures and request, AGX Thor, 2026-11-14
  • Real-world detection qualityNot yet measuredResults so far are on labeled simulation pictures
How the system fits together
  1. PhysicalCameras and sensors
  2. Edge computePerception pipeline
  3. Edge computeInference on AGX Thor
  4. Edge computeLatency and throughput telemetry
  5. Edge computeChecks against budgets
  6. PhysicalOperator review with a RAG assistant

The problem

A robot that works most of the time is not deployable. Operators need to know when perception or inference has degraded before it becomes a physical failure, and they need evidence they can review afterward.

Approach

  • Instrument the full perception and inference pipeline on Jetson AGX Thor, not just the model.
  • Measure pipeline overhead separately from model cost, using a mock model as a baseline.
  • Use retrieval-grounded agents to help an operator interpret what the telemetry shows, with a human making the call.

Where it stands

Pipeline overhead and Cosmos-Reason2-2B inference are measured on Thor. Detection quality is now measured on 60 labeled Isaac Sim pictures, not on real-world footage. The scorer and the pictures are in the full build; the summary reports are in the public repository.

  • One change to the request (2026-10-17). Making hard_hat and vest required fields took false protective-gear alarms from 11 to 0, and all 15 real ones stayed. What the model sees did not change: it still missed the person in 16 of 45 pictures and raised 16 false “too close to the robot” alarms.
  • Cosmos 3 Edge against Cosmos Reason 2 (2026-11-14). Same prompt, schema and pictures, three runs each. Cosmos 3 Edge found the person in 44 of 45 pictures against 29 and raised none of the 16 false proximity alarms, but read the gear worse: it kept 13 of 15 real gear alarms and raised 3 false ones. Run with the model card’s server command as written, about two answers in three ran to the token cap; a whitespace-free grammar setting fixed it.
  • Lighting redrawn with Cosmos Transfer (2026-11-15). Cosmos3-Nano-Sim-Transfer on the Thor redrew the access gate’s five test clips as daylight, dusk and night, at 148 to 156 s and 3.5 to 3.8 Wh a clip. With Cosmos Reason 2 reading, the gate’s decision held at 15 of 15 in every version and no worker without gear was let through; the safety-glasses read fell from 45 to 27 of 45 at night. Cosmos 3 Edge behind the same gate denied every compliant worker, so it is not a drop-in reader there.

These are 60 simulated pictures and five generated clips on one board: a signal, not an accuracy rate.

On 2026-10-29 the rules were written up as a hazard analysis in the vocabulary of SOTIF and ISO 26262: a hazard table, the triggering conditions, and the acceptance criteria any safety-related use would need. It is method and vocabulary only, with no conformance claim, and it restates the boundary: the system observes and alerts, and never stops or commands a robot.