quiet-storm/polar-tokenizer-bridge

PoLAR Tokenizer Bridge

PoLAR tokenizer checkpoint trained on BridgeData V2, presented in the paper PoLAR: Factorizing Extent and Mode in Latent Actions for Robot Policy Learning.

• Project Page: https://joon-stack.github.io/PoLAR
• Repository: https://github.com/joon-stack/PoLAR

Tokenizer Details

• Artifact: polar_tokenizer_bridge.ckpt
• Tokenizer family: hyperbolic_factorized_radprog
• Factorization: 16 radial bins + 16 direction codes
• Action-token interface: 1 radial token + 4 direction tokens

With the default 16-radius/16-direction factorization, radial IDs occupy <ACT_0> through <ACT_15>, while direction IDs occupy <ACT_16> through <ACT_31>.

This repository is part of the PoLAR code release.

Training the Tokenizer

To train the PoLAR tokenizer on BridgeData V2 RLDS format, you can use the following command from the official repository:

cd latent_action_model

torchrun --standalone --nnodes 1 --nproc-per-node 8 main_visual_vq.py fit \
    --config config/polar_tokenizer_bridge.yaml \
    --data.data_root /path/to/rlds_bridge_orig \
    --model.log_path /path/to/output/lam_bridge/logs

Citation

@misc{jeong2026polar,
  title = {PoLAR: Factorizing Extent and Mode in Latent Actions for Robot Policy Learning},
  author = {Jeong, Youngjoon and Yu, Jihwan and Jo, Minsoo and Chun, Junha and Kim, Taesup},
  year = {2026},
  eprint = {2606.21139},
  archivePrefix = {arXiv}
}