TD-MPC
Model-based reinforcement learning: a task-oriented latent dynamics model and terminal value function learned jointly by temporal-difference learning, with MPPI trajectory optimisation in latent space at each decision step.
verified 2026-10-24paper →
Architecture
Compute requirements
| Memory | Precision | Conditions | |
|---|---|---|---|
| train | ≥2 GB · ≤6 GBsrc | — | Peak training VRAM at batch size 8 with AdamW (LeRobot default), "Light BC" group (act, vqbet, tdmpc) in the LeRobot compute hardware guide; the guide groups tdmpc with the light BC policies by size. The guide calls its figures indicative, order-of-magnitude numbers for picking hardware, not exact predictions. Precision not stated. · batch 8 |
Training requirements: see also the LeRobot compute hardware guide, the authoritative source for LeRobot training envelopes.
Measured inference 0 runs · inference cost only, not task success
No measured runs yet. Submit one with its runtime, precision and camera count.
Robots running TD-MPC 1 of 24 with evidence
| Robot | State | Tier | Verified | Detail |
|---|---|---|---|---|
| UFACTORY xArm | works with caveats | T1 · upstream documented | 2026-10-24 1d ago | Evidence is the FOWM authors' real xArm 7 experiments with their own codebase; the public FOWM repo runs only the simulated xArm tasks, and LeRobot's tdmpc port documents no real-robot use. |
No verified evidence, but both are available in LeRobot (ecosystem context): AgileX PiPER, ARX5, Bimanual SO-100/SO-101, EarthRover Mini Plus, HopeJR, I2RT YAM, Koch v1.1, LeKiwi, ROBOTIS OMX, OpenArm, Reachy 2, reBot Arm B601-DM, SO-100, SO-101, Trossen WidowX AI, Unitree G1, Universal Robots UR5e, XLeRobot.
Framework support
| Framework | Support | Since | Type string | Docs |
|---|---|---|---|---|
| LeRobot | nativesrc | — | tdmpc | docs |
| LeRobot: Present at v0.3.2, the earliest tagged LeRobot release. `since` left blank: the only earlier PyPI release (0.1.0, 2024-03, a TD-MPC-centric prototype) was not inspected. Defaults: use_mpc true, horizon 5, n_action_steps 1. action_output is "other": actions come from MPPI planning over the learned model, not a direct policy head. | ||||
Checkpoints
None recorded.
Related terms
Sources
- T1 · upstream documentedpapersource 2022-03-09verified 2026-10-24 (1d ago)
TD-MPC learns a latent dynamics model and terminal value function by TD learning and plans with them.
“a framework for data-driven MPC using a task-oriented latent dynamics model and terminal value function learned jointly by temporal difference (TD) learning.”
https://arxiv.org/abs/2203.04955 - T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Policy type string is `tdmpc`.
“@PreTrainedConfig.register_subclass("tdmpc")”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/policies/tdmpc/configuration_tdmpc.py - T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Feature validation allows at most one image and requires it to be square.
“f"{self.__class__.__name__} handles at most one image for now. Got image keys {self.image_features}."”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/policies/tdmpc/configuration_tdmpc.py - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Compute hardware guide: Light BC (`act`, `vqbet`, `tdmpc`) peak VRAM ~2–6GB at BS 8 with AdamW; figures are indicative.
“| Light BC | `act`, `vqbet`, `tdmpc` | ~2–6GB | Laptop GPU (RTX 3060), L4, A10G |”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/hardware_guide.mdx