Compare robot versions. Find failures. Know what is ready for the next release. MotionBench is robot policy evaluation and regression testing for Physical AI teams.
Compare a trusted reference run with a candidate after a policy, model, hardware or environment update.
Surface trajectory, smoothness, velocity and anomaly signals when telemetry channels are present.
World-model assurance research helps teams inspect where predictions and executed outcomes diverge. Explore /world-model-assurance.
Bring evaluation into CI/CD with Motion API and MCP tools. Human review stays in the loop. Explore /robot-ci-cd.
Keep evaluation IDs, metrics, events and series as evidence for the next release conversation.
Explore a real evaluation example, or start Break the Robot for an interactive challenge. Join after the free challenge to unlock five MotionBench evaluation credits for 30 days.
MotionBench compares a reference Motion Session with a candidate and returns metrics, events and evidence. Try /break or open /benchmarks.
Keep a trusted reference run, evaluate each candidate update, and inspect where behaviour diverges. Explore /solutions/physical-ai.
Ingest recorded VLA or policy runs as robot Motion Sessions with telemetry, then compare them in MotionBench. See /benchmarks for research examples.
MotionBench is Arisole robot policy evaluation and regression testing. It compares two robot runs and shows what changed.
Convert ROS2 bags with the offline ros-bridge into Motion API sessions, then evaluate those sessions in MotionBench.
Loading Arisole. If the interactive app does not start, continue with the public links above.