BATCH: Behavior Alignment for Cross-Embodiment Learning from Humans

Zeyi Liu, Shuran Song

November 2026

TL;DR BATCH pairs human egocentric clips with robot teleop clips at the same stage of a task, then finetunes a video-action diffusion policy on the pairs with co-training and human-to-robot video style transfer. On scenes seen only in human data, real-robot performance rises by an average of 45% over co-training alone and 138% over the robot-only model.

What they do

Co-training a robot policy on human egocentric video helps, but humans and robots execute the same task at different speeds, and human clips contain segments with no robot counterpart. BATCH first finds which human frames correspond to which robot frames, without labels, and then trains a video-action policy on those aligned pairs so that scenes seen only in human data transfer to the robot.

How it works

Behavior alignment. An encoder on frozen ResNet-50 features is trained with a multi-cycle temporal cycle-consistency loss (GTCC) on head-camera video from Aria glasses, worn by the person or mounted on the robot. Subsequence DTW in that embedding space retrieves the best-matching human segment for each robot window.
Video-action model. The first 8 blocks of Wan2.1-I2V-14B predict future head-camera video and 16-step action chunks with flow matching. The model is pretrained on robot demos for 20,000 steps, then finetuned for 1,000 steps on aligned pairs.
Two finetuning objectives, alternated with probability 0.5: joint co-training, where human and robot streams share weights but not attention, and style transfer, where the robot’s future video is predicted from the aligned human clip.
Deployment. Only the robot stream runs, on ARX5 arms with a head camera and wrist webcams; inference takes 0.35-0.42 s per chunk on one RTX 5090.

Results

Alignment on out-of-domain human data: GTCC + DTW gets 57.8% phase overlap on Object in Container and 71.7% on Cup on Saucer; DINOv3 + DTW gets 42.3% and 22.1%.
Object in Container (20 rollouts, normalized score): BATCH 0.63, without alignment 0.42, EgoVerse co-training 0.40, robot only 0.28.
Cup on Saucer (30 rollouts, success rate): BATCH 83.3%, EgoVerse 63.3%, without alignment 50.0%, robot only 33.3%.

Limitations

Alignment assumes humans and robots share a behavioral structure and may struggle when strategies or action orderings differ. Alignment errors feed mismatched supervision into style transfer. Robot demonstrations are still required when finetuning with human data.

Figures

From Figure 1: BATCH performs behavior alignment between diverse out-of-domain human data and limited in-domain robot... — From Figure 1: BATCH performs behavior alignment between diverse out-of-domain human data and limited in-domain robot teleop data, accounting for differences in execution speed and irrelevant human segments. The aligned pairs are leveraged through two objectives within a downstream cross-embodiment video action model: co-training and human-to-robot style transfer. The resulting policy transfers visual knowledge from human data to robot deployment in out-of-domain scenarios.

From Figure 2: BATCH method. (A) Behavior Alignment: GTCC encoder trained with multi-cycle consistency learns shared ... — From Figure 2: BATCH method. (A) Behavior Alignment: GTCC encoder trained with multi-cycle consistency learns shared temporal embeddings from unaligned demonstrations; the trained encoder then combines with DTW to retrieve behavior-aligned human and robot segments. (B) Cross-Embodiment Video Action Model: (a) aligned human and robot trajectories are encoded as separate streams and input to a diffusion transformer; (b) the attention pattern supports both joint co-training and human-to-robot style transfer objective within the same model; and (c) at deployment, the robot stream is used to execute policy in scenarios only seen in human data.

From Figure 4: Real-world rollout results. We evaluate Object in Container across four unseen configurations includin... — From Figure 4: Real-world rollout results. We evaluate Object in Container across four unseen configurations including unseen background, target objects, and distractor objects, and Cup on Saucer across unseen objects and environments. Our method achieves the strongest overall performance.

From Figure 7: Policy performance per evaluation scenario, for Object in Container (normalized score) and Cup on Sauc... — From Figure 7: Policy performance per evaluation scenario, for Object in Container (normalized score) and Cup on Saucer (success rate).

Abstract

Human demonstrations provide a scalable source of diverse manipulation data, but leveraging them for robot learning remains challenging due to differences in embodiment, viewpoint, action spaces, and temporal structure. We introduce BATCH, a cross-embodiment learning framework that explicitly aligns human and robot behaviors. BATCH learns self-supervised temporal representations with a cycle-consistency objective to establish cross-embodiment correspondences without manual annotation. We further introduce two complementary training objectives that leverage these correspondences within an autoregressive videoaction diffusion model: behavior-aligned co-training and human-to-robot style transfer. On real-world manipulation tasks, BATCH improves generalization to unseen task configurations compared with training on unaligned data and prior cross-embodiment baselines, with additional gains from the style-transfer objective. Our results demonstrate that explicit behavior alignment enables more effective transfer from human demonstrations to robot policies.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong