Bidirectional Decoding

🌐 Website | 📄 Paper | 🤗 BiD + LeRobot | 🤖 BiD + Diffusion

Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling
International Conference on Learning Representations (ICLR) 2025
Yuejiang Liu*, Jubayer Ibn Hamid*, Annie Xie, Yoonho Lee, Maximilian Du, Chelsea Finn
IRIS Lab, Stanford University

Bidirectional Decoding (BID) samples multiple action chunks at each time step and searches for the optimal action based on two criteria:

1. backward coherence, which favors actions close to the decision made in the previous time step
2. forward contrast, which favors actions close to near-optimal plans and far from sub-optimal ones

BID improves temporal consistency over multiple time steps, while maintaining high reactivity to stochastic dynamics.

Setup

Install dependencies of the diffusion policy (approx. 20 min)

mamba env create -f conda_environment.yaml
mamba activate bid

Install additional dependencies

pip install -r requirement.txt

Download pre-trained checkpoints

gdown https://drive.google.com/drive/folders/1o8rf2Lq91D_DCq7RqZVyFAP-eMcLOAP2 -O . --folder

Download online data (optional, required only for model training)

bash script/download_dataset.sh

Decoding Scripts

The sampler folder contains a collection of test-time sampling/decoding algorithms.

• Vanilla Sampling Baseline

bash script/eval_random.sh

• Warmstart Diffusion Baseline

bash script/eval_warmstart.sh

• Temporal Ensembling Baseline (EMA)

bash script/eval_ema.sh

• Our Bidirectional Decoding (BID)

bash script/eval_bid.sh

• Dynamic object animations

bash script/animate_eval.sh

Expected Results

The separate.ipynb script summarizes representative results from different algorithms on the Pust-T task.

Deterministic (noise=0.0)	Stochastic (noise=1.0)
Method Result Vanilla (ah=1) 0.846 Vanilla (ah=8) 0.884 Warmstart (ah=8) 0.887 EMA (ah=8) 0.866 BID (ah=8) 0.928	Method Result Vanilla (ah=8) 0.582 Vanilla (ah=1) 0.805 Warmstart (ah=1) 0.852 EMA (ah=1) 0.823 BID (ah=1) 0.889

The animation folder contains sample videos of different algorithms on the Pust-T task with a dynamic block (perturb=1.0).

Vanilla Open-Loop	Vanilla Closed-Loop	EMA Closed-Loop	BID Closed-Loop

The aggregate.ipynb script summarizes the results of different algorithms on seven simulation tasks over three runs.

BID offers increasing performance gains with larger sample sizes and complements other existing methods, e.g., EMA.

Citation

If you find this code useful for your research, please consider citing our paper:

@inproceedings{liu2024bid,
  title={Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling},
  author={Yuejiang Liu and Jubayer Ibn Hamid and Annie Xie and Yoonho Lee and Max Du and Chelsea Finn},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025},
  url={https://openreview.net/forum?id=qZmn2hkuzw}
}

Acknowledgement

Our implementation is built upon the Diffusion Policy codebase