The Light CompanyFrom scene setup to a comparable result.
Prism connects the physical setup to the evaluation record. Each run starts from a defined condition and ends with an outcome your team can inspect.
Define the task and test conditions.
Choose the task, objects, policy checkpoint and success criteria. Specify the timeout, allowed interventions and reasons to reject a rollout. Save the protocol so every candidate is evaluated against the same rules.
Build a test suite from the conditions you want to cover. Start with a baseline, then add the changes that matter for your use case.
Set up the station around your workcell.
Register the camera and projector to the work surface. Check the field of view, mounting and placement accuracy against the objects and robot reach.
Prism keeps the calibration reference with the scene context. When equipment moves or the workspace changes, check the station before reusing existing conditions.
Project the target scene.
Select a saved condition or create a new one. Prism projects placement targets and labels directly onto the table. The operator positions the objects using the guidance in the workspace.
Change the condition without redrawing marks or rebuilding a fixture. The scene definition stays available for the next operator and the next run.
Verify before the rollout starts.
Cameras compare the prepared scene with the intended condition. Review missing objects, placement differences and the task's acceptance criteria before running the policy.
The record keeps the target and observation distinct. A setup rejection is handled separately from a valid trial that the policy fails.
Run the policy and capture the episode.
Your robot executes the task through its existing control system. Prism connects the episode to the scene, policy version and evaluation protocol.
Keep the recording and any intervention with the run. Use a consistent protocol for when projected guidance is visible so it does not become an unintended cue during policy evaluation.
Score, inspect and compare.
Prism scores each task step against the evaluation rubric and records the reason. The camera recording, telemetry and timestamps stay with the score, and any correction keeps its reason.
Compare candidates on the same test suite. Move from an aggregate score to task results, individual trials and the conditions behind a failure.
Replay a failure against the next checkpoint.
Select a saved condition and project the setup again. Restore the objects, verify the scene and run the new policy. Compare the episodes using the same task definition.
Build regression suites from the cases your team wants to retain. Scene history keeps those cases connected as the setup and policy evolve.
Bring Prism to your lab
Tell us about your robot, task and evaluation workflow.