Quasi-Optimal Experimental Design (QOED)

Learning What Matters: Adaptive Information-Theoretic Objectives for Robot Exploration

Install the base package. You can run Experiment 1 after this step.

conda create -n qoed python=3.12 -y
conda activate qoed
conda install conda-forge::uv -y
uv pip install -e . --no-cache

Install the MJLab and RWM dependencies. You can run Experiment 2 after this step.

uv pip install -e ".[mjlab]" --no-cache

• RSS experiments: Jackal, Franka, Go1-Quadruped
• Code optimization, e.g., torch.compile() or JAX
• Additional robot platforms, e.g., G1-Humanoid, Leap-Hand
• Additional demos, e.g., next-best-view or differentiable MPC
• All robots from MuJoCo Menagerie

qoed-jackal-demo --device cuda:0 --sweep

qoed-franka-demo --device cuda:0 --sweep

--sweep runs 10 seeds per method and prints the paper summary.

Optional: train the dynamics model and policy. Pretrained checkpoints are provided under logs/. --task selects the robot: go1, jackal, g1, or leap, and go1-rough, jackal-rough, or g1-rough for rough terrain.

python mbpo/rsl_rl/train.py --task go1 --mode pretrain --gpu-ids "[0, ]" --headless

Optional: play the pretrained policy in MuJoCo.

python mbpo/rsl_rl/play.py --task go1 --mode pretrain --gpu-ids "[0, ]" --viewer auto

Run real-world online RL, using MuJoCo as the real-system proxy.

python mbpo/rsl_rl/train.py --task go1 --mode finetune --info-gain qoed --gpu-ids "[0, ]" --viewer auto

Available info-gain modes: qoed, qoed-agnostic, boed, and nothing.

Please consider cite our work if it is interesting.

@inproceedings{yu2026qoed,
    author    = {Youwei Yu and Jionghao Wang and Zhengming Yu and Wenping Wang and Lantao Liu},
    title     = {{Learning What Matters: Adaptive Information Theoretic Objectives for Robot Exploration}},
    booktitle = {Proceedings of Robotics: Science and Systems},
    year      = {2026},
    address   = {Sydney, Australia},
    month     = {July}
}