Matrix/Policy·Reward model·Qianzhong Chen et al. (Stanford University, UC Berkeley, xdof.ai)

SARM

Stage-Aware Reward Modeling: a video-based reward model for long-horizon manipulation that jointly predicts the current task stage and within-stage progress, normalised to task progress in [0, 1]. In LeRobot it is trained on your own dataset and its progress predictions can weight policy training via Reward-Aligned Behavior Cloning (RA-BC).

verified 2026-10-24paper →

Reward models are not policies you deploy on a robot, so they are listed separately from the matrix.

Architecture

Family: reward-model
Backbone: frozen CLIP ViT-B/32 image/text encoders + transformer stage/progress headssrc
Action head: scalar-rewardsrc
Code license: Apache-2.0 (LeRobot)
Cameras: min 1 · max 1src
Language input: required

Compute requirements

No sourced requirements yet.

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.

Framework support

Framework	Support	Since	Type string	Docs
LeRobot	nativesrc	0.4.3	sarm	docs
LeRobot: A reward model, not a policy: at v0.6.1 it lives in src/lerobot/rewards/sarm and is registered with RewardModelConfig.register_subclass("sarm") (the train config has a separate `reward_model` field). Added in v0.4.3 under src/lerobot/policies/sarm; moved to src/lerobot/rewards in v0.6.0 ("Reward models refactor"). The v0.6.1 sarm.mdx training examples still pass `--policy.type=sarm`. Used to compute RA-BC sample weights (`--sample_weighting.type=rabc`), which the docs say currently supports PI0, PI0.5 and SmolVLA.

Checkpoints

None recorded.

Related terms

Reward model

Sources

T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
SARM is a video-based reward modeling framework, optionally used with RA-BC.
“SARM (Stage-Aware Reward Modeling) is a video-based reward modeling framework for long-horizon robot manipulation tasks.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/sarm.mdx
T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
SARM processor encodes images and task text with CLIP ViT-B/32 and pads robot state.
“- **Encodes** images and task text with CLIP (ViT-B/32) into `video_features` and `text_features`”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/sarm.mdx
T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
RA-BC weighting currently supports PI0, PI0.5 and SmolVLA.
“Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/sarm.mdx
T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
SARM is registered as reward-model type `sarm` (RewardModelConfig, not PreTrainedConfig).
“@RewardModelConfig.register_subclass("sarm")”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/rewards/sarm/configuration_sarm.py
T1 · upstream documentedrelease-notessource 2026-01-22verified 2026-10-24 (1d ago)lerobot 0.4.3
SARM was first added in LeRobot v0.4.3.
“* Add sarm by @pkooij in https://github.com/huggingface/lerobot/pull/2639”
https://github.com/huggingface/lerobot/releases/tag/v0.4.3
T1 · upstream documentedrelease-notessource 2026-07-06verified 2026-10-24 (1d ago)lerobot 0.6.0
v0.6.0 refactored reward models into src/lerobot/rewards.
“* Reward models refactor by @s1lent4gnt in https://github.com/huggingface/lerobot/pull/3142”
https://github.com/huggingface/lerobot/releases/tag/v0.6.0