Lucky Robots

Lucky Robots is used to build working robot skills through practical, repeatable runs in digital scenes. You start by choosing a space, selecting a robot from the catalog, and describing the job in plain English or code. Goals, constraints, and success checks are set up front, so every run produces consistent results you can compare.

A common workflow pairs perception and control. Configure the robot’s sensors, stream color and distance maps, and record rollouts while the agent explores or follows demonstrations. Those recordings become training data for policies, planners, or vision models. You can schedule many trials at once with varied layouts, lighting, clutter, and object poses to expose edge cases early.

For learning from demonstration, operators drive the robot through target behaviors and the system captures synchronized observations, actions, and rewards. For reinforcement-style training, you define rewards and termination rules, then launch batches of episodes across multiple scenes. Checkpoints are saved on a schedule, letting you compare versions on the same test suite and pick the best performer.

Debugging focuses on evidence, not guesswork. Inspect replays frame by frame, overlay detections or contact events, and jump straight to failures like collisions or timeouts. Adjust curricula, randomness, or sensor layouts and immediately see how changes affect success rates, latency, and stability. When a behavior looks solid, run a stress pass with heavy variation to verify it holds up across conditions. more