MODULE 1 / LESSON 3 · DATA, DEMONSTRATIONS & DISTRIBUTION SHIFT
Push-T: from demonstrations to a policy
Collect. Fit. Test. Use physical evidence to improve your dataset.
MEASURED RESULTS
Your experiments, side by side
Collect a dataset, train a policy and compare closed-loop outcomes.
| Run | Split | Success | Coverage |
|---|---|---|---|
| No measured runs yet. | |||
Code, equations, data contract & model limits +
Implement and reproduce
Actions are absolute world-frame pusher targets in metres, held for 50 ms. The solver runs at 240 Hz. The object moves only through physical contact.
minW ‖XW − A‖² + λ‖W‖²
X contains standardized state/goal features, A recorded two-dimensional actions, W policy weights and λ regularization. Training preprocessing never uses validation or test observations.
Learning implementation (JavaScript) · Official ACT workspace (Python) · Reproduce the fit (Python)
What this experiment can establish
Training focuses on upright translation with varied lateral starts and push distances. It is not a solution to unrestricted rotating Push-T. State is read from the simulator, not a camera or VLM. Physics parameters are instructional, not calibrated hardware.
Human, teacher, learner and correction sources are recorded separately. A learner prefix is not an expert label. The Python companion validates and replays eligible demonstrations from the JSON export before writing an official state-only LeRobotDataset. Validation, final-test and learner-prefix records are excluded from training. No RGB training data or VLA is implied.
Datasets, checkpoints and results are saved in this browser on this device. Export a backup before clearing browser storage. There is no cloud synchronization or certificate grading in this lab.