Evaluate how physical systems move.

MotionBench is Arisole’s Physical AI evaluation product. It compares a reference Motion Session with a candidate session and returns metrics, events and aligned series for how the candidate moved.

Robot run comparison

An evaluation takes two Motion Session IDs: reference_session_id and candidate_session_id. Both sessions need a robot entity. The service returns status compared, partial or cannot_compare, plus a summary of available metrics, insufficient metrics and anomalies.

Telemetry evaluation

Runs are loaded from motion observations — timestamped channels such as odometry position and joint velocities. Teams typically ingest telemetry through Motion API, including ROS2 bags converted by the offline ros-bridge into session sources.

Trajectory deviation

Trajectory evaluation inspects how candidate position series diverge from the reference after time alignment. The result is a metric with quality and confidence metadata, not a ranking of robot brands.

Smoothness

Smoothness evaluators look at how jerky or continuous a motion channel appears across the compared window.

Velocity stability

Velocity evaluation inspects stability of joint or related velocity channels between the two runs.

Repeatability

Repeatability evaluation asks how consistently the candidate traces the reference behaviour across the aligned window.

Motion anomalies

Evaluations can emit events typed as motion_deviation, anomaly or instability_signal, each with a timestamp, optional channel and severity of low, medium or high.

FAQ

What is MotionBench?

MotionBench is Arisole’s Physical AI evaluation product. It compares two robot Motion Sessions and returns metrics, events and series for how the candidate moved relative to a reference.

What data can MotionBench evaluate?

MotionBench evaluates Motion Sessions that include a robot entity and motion observations such as telemetry channels. Video can exist on a session; comparison currently uses observation channels loaded for each run.

What does MotionBench output?

An evaluation returns status, a summary, metrics (including trajectory, smoothness, velocity, repeatability and related keys when data is sufficient), events, and aligned series. Insufficient channels are marked rather than invented.

Can MotionBench compare robot runs?

Yes. POST /v1/motionbench/evaluations accepts two session UUIDs. GET /v1/motionbench/evaluations/:id retrieves a stored evaluation for the same workspace.

How does MotionBench use ROS2 telemetry?

ROS2 bags are converted outside the web API by the ros-bridge into Motion API sessions (telemetry sources, observations and events). MotionBench then compares those sessions. It does not run ROS inside the public website.

Loading Arisole. If the interactive app does not start, continue with the public links above.