This repo is anonymized in compliance with ICLR 2025 submission guidelines.
Implementation of Bootstrapped Model Predictive Control (BMPC).
Install dependencies through conda.
conda env create -f docker/environment.yaml
pip install gym==0.21.0
Depending on your existing system packages, you may need to install other dependencies. See docker/Dockerfile for a list of recommended system packages.
See below examples on how to train an BMPC agent in the default setting.
$ python train.py task=walker-walk steps=500000
$ python train.py task=dog-run steps=1000000
See config.yaml for a full list of arguments.
This codebase supports 28 tasks from DMControl, which covers all tasks used in the paper. See below table for expected name formatting:
| domain | task |
|---|---|
| dmcontrol | walker-walk |
| dmcontrol | finger-turn-easy |
| dmcontrol | cartpole-balance-sparse |
| dmcontrol | dog-run |
| dmcontrol | dog-stand |
| dmcontrol | humanoid-run |
which can be run by specifying the task argument for train.py.
We further evaluate BMPC on HumanoidBench. The corresponding evaluation performance is shown in the figure below.
In the top left, we present the average performence of all tasks except for Reach due to the different reward scales. Mean and 95% CIs over 3 seeds.
The code borrows heavily from nicklashansen's tdmpc2 implementation.