TD-MPC2
Single 317M agent trained on 104 continuous-control tasks across 4 domains, planning in a learned implicit world model. Scales predictably with model and data size.
- 85%
- 27K
- 50 Hz
- —
Benchmarks
Numbers as reported by the authors on the named evaluation suite.
| Evaluation | Suite | Score |
|---|---|---|
104-task multitask score | TD-MPC2 multitask | 70.6 |
DMControl average | DMControl | 85 |
Meta-World average | Meta-World MT50 | 75% |
Humanoid / dog locomotion | DMControl hard | 68 |
Training
Model-based RL with latent-space planning
317M
8x A100 40GB
3.4K accelerator-hours
Joint multitask training on 545M transitions with decoupled latent dynamics, reward and Q ensembles; no per-task tuning.
Same hyperparameters across all 104 tasks; capacity scaling from 1M to 317M improves the multitask score monotonically.
Linked datasets, robots & papers
Resolving related models, datasets and papers…