WorldPolicies

Robocurve's registry of published robot-policy (VLA) checkpoints.

Every checkpoint Robocurve publishes is indexed here. An entry records the weights, base model, training method, training code, and an eval number with the protocol behind it. CI validates each field, so the table below stays accurate.

The weights themselves live on the robocurve Hugging Face org and can be browsed as the WorldPolicies Collection. Each model card follows the worldevals model-card standard.

Policies

Policy	Embodiment	Base model	Method	Eval
gr00t-n1.7-so101-molmoact2	SO-101	GR00T-N1.7-3B	LoRA	eval_loss 0.0273 (held-out episodes)
gr00t-n1.7-yam-molmoact2	YAM bimanual	GR00T-N1.7-3B	full FT	open-loop MSE 0.002492 (held-out repo)
pi05-yam-molmoact2	YAM bimanual	pi05_base	full FT	open-loop MSE 0.002412 (held-out repo)
dreamzero-yam-molmoact2	YAM bimanual	DreamZero-AgiBot	LoRA r=32	open-loop MSE 0.002524 (held-out repo)

Lower is better for all metrics. The three YAM rows share a single frozen 20-sample protocol (raw absolute joints, first 16 predicted steps, cross-model ground-truth identity machine-verified; re-scored 2026-08) and are directly comparable — a statistical three-way tie, with DreamZero trained on 13-27x fewer samples. Per-sample results ship in the DreamZero release repo's eval/ folder. The SO-101 row uses a different metric and is not comparable.

Install & use

pip install "worldpolicies @ git+https://github.com/robocurve/worldpolicies"

worldpolicies list                       # all checkpoints
worldpolicies list --embodiment so101    # filter
worldpolicies info pi05-yam              # one entry in detail

from worldpolicies import catalog, get, by_embodiment

print(get("gr00t-n17-so101").hf_url)

Evaluate a policy

Serve the checkpoint (see the Usage section of its card), then drive it with Inspect Robots through the embodiment adapters (SO-101, YAM) against benchmarks from WorldEvals.

Add a checkpoint

1. Publish the model to the robocurve Hugging Face org with a card that passes the model-card standard. The checklist in the standard is gating.
2. Add it to the WorldPolicies Collection.
3. Append a Policy(...) entry to src/worldpolicies/catalog.py. CI validates every field.

Every public module, class, and function needs a docstring that states its contract rather than restating its name; Ruff D1 enforces this in CI as part of uv run ruff check ..

License

MIT for this index. Each checkpoint carries its own license, listed on its model card.