EmbodiChain models
Pretrained model bundles for EmbodiChain.
This repository contains 12 PPO policies: ANYmal-C, G1 flat locomotion, Go1, Go2,
H1_2 and MicroDuck, each with default and newton physics configurations.
- Model index: model IDs, tasks and tested runtime information.
policies/<model-id>/: weights, RunManifest, configuration snapshots and evaluation evidence.- Policy evaluation documentation: installation prerequisites, commands and options are maintained in EmbodiChain.
- Robot assets are distributed separately in
DexForceAI/embodichain_data.
HF revisions version the indexes and bundles together. Policy bundles retain their existing layout; neural dynamics comparison media lives under neural_dynamics/nerd/media/.
Evaluation and licensing
The existing PPO policy weights do not declare a weight license. EmbodiChain's code license does not establish their redistribution terms. The new neural dynamics bundles declare Apache-2.0 in their own versioned folders; that declaration does not alter the PPO policy licenses.
All 12 policies completed 1,000 control steps (20 simulated seconds) with FastRT using the public DexSim 0.5.0 wheel, Warp 1.15.0 and Newton 1.4.0. Robot assets were downloaded through the standard resolver into a new cache. G1 was also evaluated headlessly on both backends, starting with empty model and asset caches. Checkpoint training counters and runtime versions are recorded in each evaluation.json.
The evaluation used EmbodiChain main at 1b23ea0a with the proposed pretrained
download entry point. Automatic --pretrained support is being prepared for
EmbodiChain; the bundles already use its existing native RUN layout.
Neural dynamics models
Neural dynamics index lists robot-specific inference bundles under neural_dynamics/nerd/<robot>/v1/. Current Cartpole and Ant bundles include tensor-only weights, codec/physics manifests, license notices and compact evaluation summaries. Their matching HDF5 datasets are in DexForceAI/embodichain_datasets.
Use the Neural Dynamics Model and NeRD Features documentation in EmbodiChain for download and matching-scene inference. The models are research releases; Ant long-horizon drift remains significant and paper-accuracy reproduction has not been established.