Cross-Embodiment Transfer

An action vector is a control interface, not a hardware DoF count. ACT reports a 14-dimensional absolute-joint action vector for ALOHA, while its hardware section specifies two 6-DoF ViperX arms with grippers. The same paper also describes the observation coordinates as “7+7=14 DoF”; that wording should not be read as two 7-DoF arms Zhao 2023↓Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwareTony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023. Human video supplies images rather than native robot commands: EgoScale estimates hand poses, represents wrist motion relatively, and retargetsretargetingMapping motion recorded on one body, usually a human's, onto a robot with different proportions and joint limits, so human demonstrations become references the robot can track. Raw retargeting ignores the robot's physical constraints and can produce references that are morphologically infeasible, which is why systems such as H2O and ASAP pair retargeted human motion with reinforcement learning that restores balance and contact feasibility on the real body. hand articulation into a robot hand action space Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. The interface and the training recipe both matter when combining these data.

The interactive compares three accounts of cross-embodimentcross-embodimentLearning from, or transferring to, robots with different morphologies: different joint counts, action dimensions, and sensor suites. The Open X-Embodiment project pooled data from 22 robot embodiments into a single training mixture and showed that the resulting RT-X policies carry skills across platforms. The gain is not free, because each body defines its actions differently, and the field answers that mismatch with several incompatible conventions. learning with the same illustrative slot geometry. Motion Transfer is described at a high level in the Gemini report; the hatched blocks are not a recovered model architecture. Its empty human-hand row means that this diagram does not specify a hand-to-action mapping, not that the model cannot learn from human video.

The four entries use different populations and units: pi0’s own robot data Black 2024↓π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024, the OXE dataset rather than its nine-embodiment experimental mixture Open X-Embodiment Collaboration 2023↓Open X-Embodiment: Robotic Learning Datasets and RT-X ModelsOpen X-Embodiment Collaboration, Abby O'Neill, Abdul Rehman et al., 2023, NVIDIA’s rounded N1.7 pretraining report NVIDIA 2026↓NVIDIA Isaac GR00TNVIDIA, 2026.

Kareer 2025↓Emergence of Human to Robot Transfer in VLAsSimar Kareer, Karl Pertsch, James Darpinian et al., 2025 The fourth entry is PI’s four-scenario relative performance comparison. Human-video hours are not robot-demonstration hours.

Robot interfaces and an illustrative padded vector

The pi0 report describes 10,000 hours of its own dexterous-manipulation data across 7 robot configurations and 68 tasks, in addition to previously collected OXE, DROID, and Bridge data. Its mixture discussion counts timesteps: 9.1% comes from open-source datasets, while its own datasets supply 903 million timesteps Black 2024↓π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024. Those populations do not support describing most of the 10,000 hours as OXE data.

Octo describes a different interface strategy. It curates 25 OXEOpen X-EmbodimentA 2023 collaboration across 34 labs that pooled 60 existing robot datasets into one standardized format: over a million trajectories spanning 22 robot embodiments and 527 skills. The project also trained the RT-X models on the pool, and found that a policy trained across many robots outperforms versions trained on each robot's data alone, with skills carrying over to embodiments the policy saw little of. datasets using delta end-effectorend effectorThe last link of a robot arm and whatever is attached to it: the gripper, hand, or tool whose pose the arm exists to place. Kinematics is conventionally written as the map from joint angles to the end-effector pose, and most action spaces in learned manipulation are defined as end-effector deltas rather than joint targets, because a task is specified in the space where the hand meets the world. control, zero-pads missing camera channels, and aligns gripper commands so +1 means open and 0 means closed. For downstream changes to action spaces or observations, it can add lightweight encoders or output heads while retaining the pretrained transformer Octo Model Team 2024↓Octo: An Open-Source Generalist Robot PolicyOcto Model Team, Dibya Ghosh, Homer Walke et al., 2024. This is not evidence that Octo uses the article’s former widest-robot action-vector and per-embodiment normalization recipe.

The padded mode above is an original slot-layout example: it puts each toy robot’s coordinates first and fills the remainder with zeros. It does not implement either paper, and its empty human row means that this toy has no hand-to-action adapter; it does not establish that a padded model cannot learn from human data.

Moritz Reuss’s October 2025 critique says that “OXE is mostly low-quality data” and that methods for quantifying imitation-learning data quality remain inadequate. This is his stated assessment, not evidence that normalization produces a spurious transfer effect. It is also not a claim that padding cannot represent human-hand data. Reuss 2025↓State of VLA Research at ICLR 2026Moritz Reuss, 2025

Strategy two: Motion Transfer across robot embodiments

Gemini Robotics 1.5 is trained on robot data from ALOHA (citing ALOHA 2), bi-arm Franka, and Apptronik Apollo, alongside Internet text, images, and video. The v3 report describes Motion Transfer as a new architecture and training recipe for learning across embodiments, and attributes its gains to aligning embodiments and extracting shared knowledge. In the generalization ablation, the authors compare GR 1.5 against two GR 1.5 variants trained without Motion Transfer: one on a single embodiment and one on multi-embodiment data. Comparisons are interleaved on the same robot work cell; they generally report mean progress score and its standard error, not just binary success. Gemini Robotics Team 2025↓Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion TransferGemini Robotics Team, Abbas Abdolmaleki, Saminda Abeyruwan et al., 2025

The separate cross-embodiment benchmark tests tasks whose action demonstrations were collected only on another robot. The report finds that Motion Transfer helps, but its incremental effect is less pronounced for Apollo than for the other platforms. This is not evidence that every benefit of adding multi-robot data comes from Motion Transfer. The report's generalization and transfer sections use a shared checkpoint; its long-horizon agent experiment separately adds post-training for bi-arm Franka. Gemini Robotics Team 2025↓Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion TransferGemini Robotics Team, Abbas Abdolmaleki, Saminda Abeyruwan et al., 2025 Gemini Robotics On-Device 2 is described as natively multi-embodiment and inheriting Motion Transfer from 1.5. DeepMind reports adaptation to new bi-arm embodiments in a few hours of adaptation time, typically with fewer than 200 examples, and names Dexmate, SO101, and Trossen demonstrations. The announcement does not give a separate count, held-out task split, evaluation denominator, or compute-only timing protocol for each platform. This is an attributed robot-embodiment adaptation result, not evidence for arbitrary human-to-robot action mappings. Parada 2026↓Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, 2026

The disclosure is partial, not absent. The v3 model card specifies continuous numerical robot actions, and the report describes the orchestrator/VLAvision-language-action modelA robot policy built by adapting a vision-language model, pretrained on web-scale image and text data, to output robot actions. RT-2 established the approach by representing actions as text tokens and co-fine-tuning on web and robot data together, showing that semantic knowledge from web pretraining transfers into control. The same recipe underlies OpenVLA and the π series of generalist policies. split and Motion Transfer at a high level. It does not specify the internal alignment representation or a slot layout. Its discussion says the architecture can learn from human and synthetic videos without action annotations, while presenting broader use of those sources as future work. The interactive therefore leaves the specific hand-to-action mapping unmodelled; it must not turn that limitation into a claim that human video is unusable. Gemini Robotics Team 2025↓Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion TransferGemini Robotics Team, Abbas Abdolmaleki, Saminda Abeyruwan et al., 2025

Strategy three: one relative end-effector space for robots and humans

NVIDIA’s N1.7 README describes a relative end-effector action space shared across human and robot embodiments, with actions expressed as deltas from the current pose. It calls this representation a key factor in cross-embodiment performance and reports 20K hours of EgoScale human video in pretraining alongside diverse robot demonstrations NVIDIA 2026↓NVIDIA Isaac GR00TNVIDIA, 2026. That README statement does not establish the former claim that EmbodimentTag is a geometric frame transform.

The separate EgoScale v1 paper makes the motion representation more specific. It uses relative wrist motion and retargets 21 human hand keypoints into a 22-DoF Sharpa hand action space. Robot arms use relative end-effector control, but the hands use target joint angles. Its embodiment-conditioned MLP adapters encode proprioceptionproprioceptionSensing of the body's own state: joint positions and velocities, orientation, and contact, as opposed to exteroception, which senses the outside world through cameras or lidar. Proprioception is fast, cheap, and never occluded, and the ETH Zurich line showed how far it goes: Lee and colleagues trained an ANYmal controller that hiked mud, snow, rubble, and vegetation on proprioception alone, and Miki and colleagues kept it as the trusted channel the policy falls back on when its terrain map disagrees with its body. and decode hand actions; wrist prediction, the backbone, and the action expert are shared Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. EgoScale also uses aligned human-robot mid-training, so its method should not be described as transfer with “no domain-adaptation step.” The N1.7 README’s pretraining statement and the EgoScale paper’s full training pipeline are distinct claims.

Two reported human-to-robot transfer recipes

Physical Intelligence’s Simar Kareer and colleagues reported on December 16, 2025 that they co-fine-tuned pretrained pi0.5 on human data and the most relevant robot data. Human demonstrations were treated as another embodiment, with 3D hand positions as actions, without a special transfer-learning mechanism. The note reports about 2× performance across four generalization scenarios present only in the human demonstrations: bussing, spice, dresser, and eggs Kareer 2025↓Emergence of Human to Robot Transfer in VLAsSimar Kareer, Karl Pertsch, James Darpinian et al., 2025. This is a relative performance ratio, not a two-percentage-point increase or a result over arbitrary robot tasks. The note does not give a numeric evaluation-trial denominator for this aggregate.

Their feature analysis projects mean-pooled final-layer VLM tokens with t-SNE. Greater human-robot feature overlap correlates with the generalization results as robot pretraining becomes more diverse; human fine-tuning data is held fixed between these models Kareer 2025↓Emergence of Human to Robot Transfer in VLAsSimar Kareer, Karl Pertsch, James Darpinian et al., 2025. This is the reported analysis, not a proved universal scale threshold for alignment.

Ruijie Zheng and the EgoScale coauthors take a more explicitly aligned route for dexterous manipulation. They write that effective transfer in their setting requires scale, explicit motion supervision, and a small amount of precise human-robot alignment. They also report that large-scale human pretraining without precise embodiment alignment already outperforms their mid-training-only baseline across most tasks; adding aligned mid-training gives their best overall performance Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026.

Keep both results scoped to their experiments. PI studies robot pretraining followed by human-robot co-fine-tuning on four generalization scenarios Kareer 2025↓Emergence of Human to Robot Transfer in VLAsSimar Kareer, Karl Pertsch, James Darpinian et al., 2025; EgoScale studies large-scale human pretraining, aligned mid-training, and task-specific robot post-training Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. These two sources do not establish a winner at matched data, hardware, compute, and tasks. No claim about the existence or absence of other head-to-head experiments is made here.