We open-source a total of 324 TD-MPC2 model checkpoints, including 12 multi-task models (ranging from 1M to 317M parameters) trained on 80, 70, and 30 tasks, respectively. We are excited to see what the community will do with these models, and hope that our release will encourage other research labs to open-source their checkpoints as well.

Multi-task Checkpoints

We recommend starting out with the 48M checkpoints, which strike a good balance between model size and performance. The 317M checkpoints are the largest models we trained, and achieve the best performance on both task sets. However, they are also the most expensive to train and evaluate. Note that "Score" is not comparable across task sets.

Tasks Domains Params Score Link
80 DMControl + Meta-World 1M 15.98 Download
80 DMControl + Meta-World 5M 49.45 Download
80 DMControl + Meta-World 19M 57.13 Download
80 DMControl + Meta-World 48M 68.03 Download
80 DMControl + Meta-World 317M 70.63 Download
70 DMControl + Meta-World 5M 49.3 Download
70 DMControl + Meta-World 19M 67.0 Download
30 DMControl 1M 18.93 Download
30 DMControl 5M 28.32 Download
30 DMControl 19M 54.22 Download
30 DMControl 48M 59.43 Download
30 DMControl 317M 71.41 Download

Single-task Checkpoints

We organize single-task checkpoints by task domain, totalling 312 checkpoints for 104 tasks across 4 domains. Most, but not all, models were trained to convergence and produce expert behavior.

Tasks Domain Params Checkpoints Total size Link
39 DMControl 5M 117 3.6GB Download
50 Meta-World 5M 150 4.9GB Download
5 ManiSkill2 5M 15 505MB Download
10 MyoSuite 5M 30 1.0GB Download

We also provide raw per-task and per-seed scores for all 104 tasks. They can be downloaded here.