A foundation model for humanoids, trained with no task rewards

What it does

BFM-Zero is a single neural policy pretrained on humanoid motion via unsupervised reinforcement learning. After training, you extract a latent vector z from a motion clip, a goal pose, or a reward specification to prompt the policy into tracking, reaching, or reward-based behavior without retraining. It exports policies to ONNX for sim-to-sim or sim-to-real deployment.

The interesting bit

The project treats robot control like a foundation model problem: pretrain once without task labels, then prompt at inference time. The unsupervised phase learns a general behavioral prior from the LaFan motion dataset, and the latent prompt space acts as a compact interface for downstream tasks.

Key highlights

• One unsupervised training run produces a policy that handles motion tracking, goal reaching, and reward tasks via different latent prompts.
• Ships with pretrained checkpoints and three inference pipelines that export directly to ONNX.
• Supports both Isaac Sim and MuJoCo; the MuJoCo path runs without Isaac Lab for lighter sim-to-sim testing.
• Licensed under CC BY-NC 4.0, so it is free for research but locked out of commercial use.

Caveats

• Minimal training code targeting an RTX 4090 is still listed as unreleased.
• The repository relies on Git LFS for motion data, and the README warns that LFS bandwidth limits may force you to fetch assets from HuggingFace instead.
• Training to convergence is heavy: the authors expect 50–100 million environment steps and a CUDA-capable GPU.

Verdict

Worth a look if you are doing humanoid sim-to-real research and want a pretrained unsupervised prior to prompt or fine-tune. Skip it if you need a commercial license or are hoping for lightweight, consumer-hardware training out of the box.