Training and Evaluation

Training goes through the top-level pixi tasks. Pass the task ID with --task.

Train

pixi run train --task WujiHand_Reorient --agent.upload-model False

--agent.upload-model False keeps checkpoints local only. Drop it and set WANDB_API_KEY to also push the final-iteration checkpoint to W&B as a model artifact. Either way, local .pt files land on every save_interval boundary.

Checkpoints and W&B logs write to logs/rsl_rl/<run_name>/.

Task Configurations

Two task IDs share the same MDP but trade GPU memory against policy quality.

Task ID	Envs × Iterations	GPU memory	Notes
`WujiHand_Reorient`	8192 × 5000	~20 GB	Release config, reproduces the released checkpoint
`WujiHand_Reorient_Light`	4096 × 7500	~12 GB	Lower-VRAM variant, visibly weaker policy

If pixi run train runs out of memory, switch to the lower-VRAM variant:

pixi run train --task WujiHand_Reorient_Light

WujiHand_Reorient_Light fits comfortably under ~12 GB but converges to a weaker policy, with occasional cube drops and finger-jam behavior on harder reorientations.

Replay and Evaluate

Replay a trained checkpoint in the interactive viewer, or run a success-rate eval over many trials:

# Interactive viewer with a trained checkpoint
pixi run play --task WujiHand_Reorient --checkpoint-file <path-to-ckpt.pt>

# Success-rate eval over N trials (consumes ONNX)
pixi run python -m wuji_mjlab.tasks.reorient.scripts.eval_success_rate <path-to-policy.onnx>

For headless, machine-readable output suited to CI or sweeps:

pixi run python -m wuji_mjlab.tasks.reorient.scripts.eval_success_rate <onnx_path> \
    --num-trials 100 --no-viewer --json-output result.json

Programmatic Evaluation

The eval core is importable, so you can drive batch evaluation from Python without the CLI:

from pathlib import Path
from wuji_mjlab.tasks.reorient.tooling.eval_core import EvalConfig, run_eval

result = run_eval(EvalConfig(
    onnx_path=Path("<path-to-policy.onnx>"),
    num_trials=50,
    no_viewer=True,
))
print(f"success_rate = {result.success_rate:.2%}")
print(f"mean min ori error = {result.mean_min_ori_error_rad:.3f} rad")

# Per-trial data for custom analysis
for trial in result.trials:
    if trial.status == "success":
        print(f"trial {trial.trial_idx}: t_first_succ={trial.time_to_first_success_s:.2f}s")

Export to ONNX

Deployment consumes an ONNX policy plus a sidecar JSON that records the control-mode parameters (action_scale, ema_alpha, ctrl_dt) captured at export time:

pixi run python -m wuji_mjlab.tasks.reorient.scripts.export_onnx <path-to-ckpt.pt>

The sidecar keeps deploy inference identical to the sim policy that produced the ONNX. Continue to Sim-to-real Deployment to run it on the physical hand.

Development Utilities

# List all registered task IDs
pixi run list-envs

# View the task scene with a dummy policy
pixi run python -m wuji_mjlab.tasks.reorient.scripts.view_task WujiHand_Reorient