Independent product concept · Physical AI safety layer
Antibody
A second thought before a robot moves.
Antibody is a safety layer between a robot’s AI and its motors. It checks actions against the scene’s rules, watches for changes, and records what happened.
Being tested in simulationWhy it matters
A safe request can become an unsafe move.
The camera sees one moment. A person reaches in the next. Even a planner that refused harmful requests made unsafe moves in these tests. Antibody keeps checking as the scene changes.
Part of a wider safety stack.
A broad robotics safety stack that Antibody is designed to complement.
A harmful-instruction benchmark. Its scenarios are adapted for these simulation tests.
Scene rules, action checks, motion monitoring, and evidence that can be replayed.
Simulation results
What the simulation shows.
The guard reduced unsafe outcomes in fixed tests. Sudden reaches still exposed a limit when camera updates were delayed.
unsafe outcomes across 44 paired kitchen tests
useful tasks completed with the guard, for each of four planners
randomized runs still unsafe when camera updates came too late
Four planners, the same 11 kitchen scenes, each run with and without Antibody. The randomized set used a different mix of changing scenes. Simulation only; no hardware validation.
Test conditions and limits
- The fixed suite used one scripted planner, Claude, and two Qwen models.
- Unsafe outcomes follow project rules for contact, proximity, heat, chemicals, and other known hazards.
- The kitchen simulation waits for the guard at each step and models 0.25 seconds of camera delay. It ran 4.2 times slower than real time, so these tests do not establish live robot timing.
- In the randomized set, 31 of 58 useful tasks finished. This is separate from the fixed suite’s 5 of 6 result.
- Scenes use scripted people, clean depth images, and simplified object handling. RoboHarm tasks use local reconstructions and project scoring rules.
The product opportunity
Keep the evidence when the AI changes.
Saving failures as tests helps teams check later model and policy changes. The potential advantage is a growing library of scene rules, repeatable scenarios, and comparable evidence.
- 01
Run the scene
Apply the workspace’s rules to the proposed action.
- 02
Keep the evidence
Record the observations, decision, and outcome.
- 03
Retest the change
Reuse the failure case when the AI or policy changes.
Next: Antibody Test. Package this workflow so a team can test its robot, AI, and safety rules in a specific scene, then review a replay and signed report.
Recorded examples
See what gets caught.
Short recorded clips from the test bench, including the original run behind the opening visual.
The recorded reach, side by side
A person reaches into the arm’s path. The unguarded run continues; the guarded run slows and stops.
Open clipTools & resources
Inside the prototype.
In the kitchen tests, the guard reads camera and robot messages without access to hidden scene state. The humanoid demo uses simulated position and speed as stand-ins for robot odometry.
- Isaac Sim + OpenUSD
- Robot scenes, physics, and simulated cameras built on NVIDIA Omniverse.
- ROS 2
- Connects the planner, guard, controller, and recorder through robot interfaces.
- MCAP + signed records
- Keeps sensor and decision history, with signatures that reveal later edits.
- NVIDIA Cosmos
- Creates viewing footage from recorded simulations, including the opening clip.