3.1 KiB
Train and replay on Colab
Upload a copy of notebooks/robot_colab.ipynb to Google Colab. Select
Runtime → Change runtime type → GPU. Keep the notebook and source checkout
on matching revisions.
Prepare a new robot run
- Onboard the robot locally using the workflow. Commit its
literal
tasks/<robot_id>/robot_spec.py, MJCF, referenced assets and any control file to the GitHub revision you intend to clone. Keep the model's asset paths portable within that checkout. - Fill
TASKwith the robot ID andMODELwith its registered model name frompython scripts/train.py --list. The notebook requires both explicitly. In Advanced source enterREPOSITORY_URLandSOURCE_REVISIONfor the checkout containing those inputs. - Run setup. It verifies the requested commit, isolated interpreter, CUDA arithmetic, Warp kernel and NVIDIA EGL before training. Reused checkouts are never switched or reset automatically.
- Choose
NUM_ENVSandITERATIONSfor the model and GPU. The optional smoke uses at most 256 environments for five updates. Optional preview takes a compatibleDEMO_CHECKPOINTfor this task and model. Both run separately from the selected checkpoint.
Resume a recorded run
Select Continue training. Upload a backup ZIP or choose a completed Drive checkpoint snapshot. The notebook reads its bounded metadata before installation and restores the recorded task, model, batch, commit, and core package versions. Older backups may require the original repository URL in Advanced source. Use the notebook from the recorded source if its workflow differs.
A restore validates file paths, archive checksums, and checkpoint hashes
before choosing the recorded checkpoint. Resume also checks the source
identity and runtime settings. ITERATIONS is the additional update budget.
The optimizer and curriculum state resume with the policy. Environments
begin new episodes. Re-running a settings cell keeps the selected checkpoint.
Evaluate and keep results
Training writes logs/<model>/<robot_id>/<timestamp>/. The progress cell
reads the selected run's TensorBoard events and exports curves and scalar
data. Replay can record inline video and joint measurements. Velocity replay
uses the checkpoint's command ranges. The fixed-command cell tests forward,
sideways, turn, and stand separately. It skips axes fixed at zero, such as
wheeled-biped lateral motion. Set custom commands within the trained ranges.
Report the scene, noise setting, command, tracking errors, duration, and resets.
The export cell writes the ONNX graph, layout, checkpoint copy, and README. Hybrid wheeled policies export with the native layout and replay model. A failed replay or export does not prevent backup of a complete checkpoint. Download the run ZIP or verify the optional Drive snapshots before the runtime ends. Saving the notebook itself preserves code and displayed outputs.
Colab files, recorded simulation and adapter parity describe the selected model and runtime. Use native command verification before packaging a controlled application.