Physical AI needs more than a demo clip. Teams compare a reference run with a candidate run and inspect metrics that stay honest when data is missing. MotionBench is that evaluation layer for robot and machine motion.
Ingest each robot trial as a Motion Session with a robot entity and telemetry sources (CSV, JSON, or ROS2 bags via ros-bridge). Call POST /v1/motionbench/evaluations with the reference and candidate session IDs. Inspect metrics, anomaly events and series. Iterate the policy, then evaluate again against the same reference.
MotionBench for evaluation. Arisole Fusion for multimodal context. Motion API for sessions and ingest. StrideIQ is the human product and is not a robot controller.
Keep a frozen reference bag and fail the evaluation when trajectory deviation or anomalies jump.
Compare walking controllers on smoothness and velocity stability instead of watching two videos side by side.
Treat partial and cannot_compare statuses as real outcomes when a channel is missing.
Share evaluation IDs so research and controls look at the same JSON, not a private notebook.
No. MotionBench evaluates recorded or ingested motion. It is not a fleet operations product.
Yes, by converting bags with the offline ros-bridge into Motion API sessions, then evaluating those sessions in MotionBench.
Read Motion API for session and source contracts, then MotionBench for evaluation routes.
Loading Arisole. If the interactive app does not start, continue with the public links above.