TD-MPC2
- AI model
- Humanoids and robot learning
TD-MPC2 is an official research implementation of a scalable model-based reinforcement learning algorithm for continuous control, trained and evaluated on single or multiple tasks.
Source on GitHub: github.com/nicklashansen/tdmpc2 (opens in a new tab). Facts read at commit e9f5932 (opens in a new tab).
Licenses
- Code
- MIT (opens in a new tab)
Read by Robotheca from LICENSE (opens in a new tab) (MIT) at commit e9f5932 (opens in a new tab), on 16 Oct 2026.
- Commercial use
- Yes
What it does
- Trains and evaluates single-task online RL and multi-task offline RL agents for continuous control.
- Covers 104 continuous control tasks across DMControl, Meta-World, ManiSkill2 and MyoSuite.
- Provides 300+ pretrained model checkpoints, including 12 multi-task models, plus 30-task and 80-task datasets.
- Supports both state and pixel (rgb) observations.
- Adds optional episodic RL support for tasks with terminations, disabled by default for reproducibility.
- Training and evaluation are run with evaluate.py and train.py, configured through config.yaml.
Written by Robotheca from the project's README (opens in a new tab) at commit e9f5932.
Getting started
- Difficulty
- Advanced
- Hardware it needs, as the project states it
- A machine with a GPU and at least 12 GB of RAM for single-task online RL; 128 GB of RAM for multi-task offline RL on the 80-task dataset. A GPU with at least 8 GB of memory is recommended for single-task online RL and evaluating the multi-task models (up to 317M parameters); training the 317M parame
- How to start
- The project's website (opens in a new tab)
Maintenance
- Status
- Maintained: a commit in the last 12 months
- Newest commit
- 14 Jul 2026 on main
- Latest release
- None published on GitHub
Checked on GitHub on 16 Oct 2026.
Links
Robots on Robotheca that use it
No robot on Robotheca is linked to it yet. Links come from a robot's robot.yaml, from its dependency files read as text, or from Robotheca's editors.
Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.