Chaoqi-LIU / oat

What is oat?

oat is a Python repository implementing Ordered Action Tokenization (OAT) for robotics RL, converting continuous action sequences into discrete, ordered tokens that transformer policies predict autoregressively. It tackles long-horizon planning in sim environments like LIBERO by enabling efficient training of tokenizers and policies, with scripts for dataset conversion from HDF5 to Zarr, multi-GPU tokenizer/policy training via Accelerate, and sim evaluation. Users get plug-and-play baselines outperforming diffusion on benchmarks, complete with Hydra configs and WandB logging.

Why is it gaining traction?

oat stands out with SOTA success rates on LIBERO tasks via OAT's register-based encoding, beating binary tokenizers and diffusion while using less compute. The quick-start workflow—clone with submodules, uv sync or conda env, run dataset scripts, then train/eval—hooks devs tired of custom RL pipelines. Modular swaps for FAST/QueST tokenizers and policies make experimentation fast, unlike rigid diffusion code github rl setups.

Who should use this?

RL researchers benchmarking imitation learning on LIBERO/robosuite manipulation tasks. Robotics engineers needing discrete action models for transformer policies over continuous diffusion. Devs replicating the OAT paper or prototyping oat rl baselines before real-robot transfer.

Verdict

Grab it for LIBERO experiments—clear README, eval scripts, and 24 stars make setup straightforward despite 1.0% credibility score signaling early maturity. Polish docs and add tests to boost adoption; solid for research now.

(198 words)