Data QA
Sensor alignment · Label consistency · Episode completeness
01 / Find the failures
Audit training episodes and compare physical trials with simulation. Connect sensor evidence to failed actions.
Perception
Object identity · Depth uncertainty · Occlusion
Actions & policies
Action timing · Physical constraints · Task success
Environments
Scene variation · Reset conditions · Reward criteria
02 / Simulate the conditions
Scroll to explore A clear view is only the beginning.
Define the workspace and task. Align camera views with sensor observations.
- Scene configuration
- Sensor alignment
Illustrative scenario
Move beyond the clean tabletop.
Vary clutter, occlusion, and object properties to expose failures in simulation.
- Simulation environments
- Perception & contact
Illustrative scenario
Follow the attempt, not just the score.
Inspect failed attempts to plan the next physical trial or simulation.
- Episode analysis
- Evaluation & RL
Illustrative scenario
03 / Run RL loops
Design RL tasks in simulation with configurable scenes, reset conditions, and task-specific rewards.
01
Build tasks from failed episodes
02
Run simulation rollouts across scene variations
03
Use outcomes to improve the policy
04 / Validate the policy
Track known failure conditions across policy versions. Validate gains in new physical trials.
01
Rerun a consistent simulation suite
02
Compare task success and constraint violations
03
Validate transfer with physical trials