REMAC-Kinetix

The simulation experiments' source code and implementation of paper Real-Time Robot Execution with Masked Action Chunking.

Installation

# Clone Kinetix submodule
git submodule update --init

# Install uv
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install dependencies
uv sync

Pipeline

1. First follow the implementation of RTC to produce the base checkpoints:

# Train expert policies with RL. Checkpoints, videos, and stats are written to ./logs-expert/<wandb-run-name>
uv run src_lora/train_expert.py 

# Generate data using experts. Data is written back to `./logs-expert/<wandb-run-name>/data/`
uv run src_lora/generate_data.py --config.run-path ./logs-expert/<wandb-run-name>

# Train imitation learning policies
uv run src_lora/train_flow_base.py --config.run-path ./logs-expert/<wandb-run-name>

# Evaluate imitation learning policies if you want
uv run src_lora/eval_flow_no_lora.py --config.run-path ./logs-bc/<wandb-run-name> --output-dir <output-dir>

2. Change the <wandb-run-name> to base_model in logs-bc. Then finetune the trained policies with:

bash run_all.sh

The above script will train and evaluate the 12 experiments in sequence.

Note: As there are 12 tasks, the number of tasks running should be divisible by the number of GPUs you use.

Acknowledgement

Thanks to these amazing repositories: RTC, openpi and other inspiring works.

Citation

If you find this work useful, please consider citing:

@misc{wang2026realtimerobotexecutionmasked,
      title={Real-Time Robot Execution with Masked Action Chunking}, 
      author={Haoxuan Wang and Gengyu Zhang and Yan Yan and Yuzhang Shang and Ramana Rao Kompella and Gaowen Liu},
      year={2026},
      eprint={2601.20130},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2601.20130}, 
}