Cross-Embodiment Transfer

Cross-embodiment transfercross-embodimentCross-embodiment learning is learning from, or transferring to, robots with different morphologies: joint counts, action dimensions and sensor suites. The Open X-Embodiment project pooled data from 22 robot embodiments, and its RT-X policies carried skills across platforms. Each body defines actions differently, and the field aligns them with several incompatible conventions. is training one policy on data from bodies with different joint counts, action dimensions and sensors, or moving a policy to an unseen body. Open X-Embodiment pools data from 22 robot embodiments^{Source 3}3 Open X-Embodiment Collaboration et al. 2023, Open X-Embodiment: Robotic Learning Datasets and RT-X Models. Open X-Embodiment Collaboration, Abby O'Neill, Abdul Rehman et al., 2023, and human video adds a body whose actions are estimated from images^{Source 8}8 Zheng et al. 2026, EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data. Ruijie Zheng, Dantong Niu, Yuqi Xie et al., 2026.

Paper figure: Open X-Embodiment Collaboration / arXiv. Licence: CC BY 4.0.

Try “Translate between bodies”.

Counting robot and human data

π0's 10,000 hours are its own data from 7 robot configurations and 68 tasks; by timesteps, the open-source OXE, DROID and Bridge data are 9.1% of the mixture^{Source 1}1 Black et al. 2024, π0: A Vision-Language-Action Flow Model for General Robot Control. Kevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024. RT-X experiments used 9 of OXE's 22 embodiments^{Source 3}3 Open X-Embodiment Collaboration et al. 2023, Open X-Embodiment: Robotic Learning Datasets and RT-X Models. Open X-Embodiment Collaboration, Abby O'Neill, Abdul Rehman et al., 2023, and GR00T N1.7's rounded 20K pretraining hours are EgoScale human video^{Source 7}7 NVIDIA 2026, NVIDIA Isaac GR00T. NVIDIA, 2026.

Moritz Reuss wrote in October 2025 that "OXE is mostly low-quality data" and that imitation learning lacks good ways to measure data quality^{Source 4}4 Reuss 2025, State of VLA Research at ICLR 2026. Moritz Reuss, 2025. Re-Mix's learned mixture weights beat uniform weights by 38% on average^{Source 11}11 Hejna et al. 2024, Re-Mix: Optimizing Data Mixtures for Large Scale Imitation Learning. Joey Hejna, Chethan Bhateja, Yichen Jiang et al., 2024, and Toyota Research Institute's controlled study of 89 policies found co-training with cross-embodiment robot data improves generalization to distribution shifts and unseen tasks^{Source 28}28 Lin et al. 2026, A Systematic Study of Data Modalities and Strategies for Co-training Large Behavior Models for Robot Manipulation. Fanqi Lin, Kushal Arora, Jean Mercat et al., 2026.

Robot action interfaces

π0 zero-pads every action vector to the largest robot's 18 dimensions^{Source 1}1 Black et al. 2024, π0: A Vision-Language-Action Flow Model for General Robot Control. Kevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024. Octo curates 25 OXEOpen X-EmbodimentOpen X-Embodiment is a 2023 collaboration of 34 labs that pooled 60 existing robot datasets into one standardized format: over a million trajectories from 22 embodiments, covering 527 skills. Its RT-X models, trained on the pool, beat policies trained on each robot's data alone and carried skills to embodiments they saw little of. datasets with delta end-effectorend effectorAn end effector is the last link of a robot arm plus its gripper, hand or tool, whose pose the arm exists to place. Kinematics is conventionally the map from joint angles to that pose. Many learned manipulation policies act in end-effector deltas, others in joint targets; across 13,000+ real rollouts, joint space favored control stability and task space favored generalization. control and aligns grippers to one convention, adding lightweight encoders or heads for new spaces^{Source 2}2 Octo Model Team et al. 2024, Octo: An Open-Source Generalist Robot Policy. Octo Model Team, Dibya Ghosh, Homer Walke et al., 2024, while CrossFormer trains one transformer on 900K trajectories from 20 embodiments without manually aligning observation or action spaces^{Source 12}12 Doshi et al. 2024, Scaling Cross-Embodied Learning: One Policy for Manipulation, Navigation, Locomotion and Aviation. Ria Doshi, Homer Walke, Oier Mees et al., 2024.

HPT maps each embodiment's proprioception and vision to a few tokens for a shared trunk, scaling to 52 datasets^{Source 13}13 Wang et al. 2024, Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers. Lirui Wang, Xinlei Chen, Jialiang Zhao et al., NeurIPS 2024. LAP writes low-level actions as natural language, and LAP-3B reports over 50% average zero-shot success on unseen robots, about twice the strongest prior VLAs^{Source 14}14 Zha et al. 2026, LAP: Language-Action Pre-Training Enables Zero-shot Cross-Embodiment Transfer. Lihan Zha, Asher J. Hancock, Mingtong Zhang et al., 2026.

Motion Transfer across robot embodiments

Gemini Robotics 1.5 trains on ALOHA, bi-arm Franka and Apptronik Apollo data plus Internet text, images and video, with Motion Transfer, an architecture and training recipe that aligns embodiments. On tasks demonstrated only on another robot it helps, less so for Apollo, in ablations run interleaved on one work cell; one checkpoint drives all three robots, though the long-horizon agent adds bi-arm Franka post-training, and the report specifies an orchestrator over the VLAvision-language-action modelA vision-language-action model is a vision-language model, pretrained on web-scale image and text data, adapted to output robot actions. RT-2 established it by encoding actions as text tokens and co-fine-tuning on web and robot data, showing that web-learned semantic knowledge transfers into control. OpenVLA and the π series of generalist policies use the same recipe. but not the alignment representation^{Source 5}5 Gemini Robotics Team et al. 2025, Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer. Gemini Robotics Team, Abbas Abdolmaleki, Saminda Abeyruwan et al., 2025. DeepMind says tasks shown only to ALOHA 2 "just work" on Apollo and bi-arm Franka, and vice versa^{Source 15}15 Parada 2025, Gemini Robotics 1.5 brings AI agents into the physical world. Carolina Parada, Google DeepMind blog, 2025.

DeepMind says On-Device 2, natively multi-embodiment with Motion Transfer, adapts to new bi-arm robots in a few hours, typically from fewer than 200 examples^{Source 6}6 Parada 2026, Gemini Robotics 2 brings whole body intelligence to robots. Carolina Parada, 2026.

Relative end-effector actions for robots and humans

GR00T N1 embedded state and action with embodiment-aware encoders^{Source 16}16 NVIDIA et al. 2025, GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. NVIDIA, Johan Bjorck, Fernando Castañeda et al., 2025; N1.6 moved most embodiments to state-relative action chunks, which NVIDIA found smoother and more accurate than absolute actions^{Source 17}17 GEAR Team et al. 2025, GR00T N1.6: An Improved Open Foundation Model for Generalist Humanoid Robots. GEAR Team, Allison Azzolini, Johan Bjorck et al., 2025. N1.7 shares relative end-effector deltas between human and robot embodiments and calls this a key factor in cross-embodiment performance^{Source 7}7 NVIDIA 2026, NVIDIA Isaac GR00T. NVIDIA, 2026.

Paper figure: Ruijie Zheng et al. / arXiv. Licence: CC BY 4.0.

EgoScale uses relative wrist motion and retargetsretargetingRetargeting is mapping 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 morphologically infeasible references. H2O and ASAP therefore add reinforcement learning that restores balance and contact feasibility on the real body. 21 human hand keypoints to the 22-DoF Sharpa hand, with embodiment-conditioned adapters for proprioceptionproprioceptionProprioception is sensing the body's own state (joint positions and velocities, orientation, contact), as opposed to exteroception of the outside world through cameras or lidar. It is fast, cheap and never occluded. At ETH Zurich, Lee et al.'s ANYmal controller hiked mud, snow, rubble and vegetation on proprioception alone. Miki et al. kept it as the trusted fallback when the terrain map disagrees with the body. and hand actions around a shared backbone and action expert^{Source 8}8 Zheng et al. 2026, EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data. Ruijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. Morphometric Imitation (September 2026) retargets reconstructed hand-object interactions across hand morphologies, then refines them with residual RL; on a Sharpa hand its policies reached 89.3% zero-shot success in 300 real trials on 30 objects^{Source 10}10 Sadjadpour et al. 2026, Morphometric Imitation: From Morphology and Contact Aware Hand Retargeting to Sim-to-Real Visuomotor Policy. Tara Sadjadpour, Siming He, C.K. Wolfe et al., 2026.

Closing the visual embodiment gap

Mirage masks out an unseen target robot and inpaints the source robot in real time^{Source 18}18 Chen et al. 2024, Mirage: Cross-Embodiment Zero-Shot Policy Transfer with Cross-Painting. Lawrence Yunliang Chen, Kush Hari, Karthik Dharmarajan et al., RSS 2024. RoVi-Aug synthesizes training images with other robots and viewpoints and deploys zero-shot on an unseen robot, and co-training with it raises success by up to 30%^{Source 19}19 Chen et al. 2024, RoVi-Aug: Robot and Viewpoint Augmentation for Cross-Embodiment Robot Learning. Lawrence Yunliang Chen, Chenfeng Xu, Karthik Dharmarajan et al., CoRL 2024.

Latent actions as a shared interface

LAPA learns discrete latent actions between video frames, pretrains a VLA on them and fine-tunes on robot data; training only on human video still transfers positively^{Source 20}20 Ye et al. 2025, Latent Action Pretraining from Videos. Seonghyeon Ye, Joel Jang, Byeongguk Jeon et al., ICLR 2025. LAC-WM (Stanford and Meta FAIR, October 2026) conditions a robot world model on a latent action space shared across embodiments; picking action chunks for a VLA on five unseen LIBERO tasks, it raised average success from 82.4% with an explicit-action world model to 92.0%^{Source 32}32 Huang et al. 2026, Cross-Embodiment Robot Foundation World Models with Latent Actions. Huang Huang, Sriram Yenamandra, Arjun Majumdar et al., 2026.

Human-to-robot transfer recipes

Simar Kareer's EgoMimic treated human and robot data equally and found an extra hour of hand data worth more than one of robot data^{Source 21}21 Kareer et al. 2024, EgoMimic: Scaling Imitation Learning via Egocentric Video. Simar Kareer, Dhruv Patel, Ryan Punamiya et al., 2024. On December 16, 2025, Kareer and Physical Intelligence colleagues co-fine-tuned π0.5 on human data, with 3D hand positions as actions, and the most relevant robot data. PI reports about 2× performance on four scenarios present only in human data, without trial counts, and transfer that improves as robot pretraining diversifies^{Source 9}9 Kareer et al. 2025, Emergence of Human to Robot Transfer in VLAs. Simar Kareer, Karl Pertsch, James Darpinian et al., 2025.

Paper figure: Marion Lepert et al. / arXiv. Licence: CC BY 4.0.

EgoScale's Ruijie Zheng and coauthors instead pretrain on human video, mid-train on aligned human-robot data and post-train on task robot data, and find adding aligned mid-training best overall^{Source 8}8 Zheng et al. 2026, EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data. Ruijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. EgoBridge aligns human and robot policy latents with optimal transport and reports a 44% absolute success gain over human-augmented cross-embodiment baselines^{Source 22}22 Punamiya et al. 2025, EgoBridge: Domain Adaptation for Generalizable Imitation from Egocentric Human Data. Ryan Punamiya, Dhruv Patel, Patcharapong Aphiwetsa et al., NeurIPS 2025. EgoLAP (Princeton, Toyota Research Institute and Physical Intelligence, October 2026) routes human and robot trajectories through a shared chain of language actions and reached 80.1% mean real-world task progress on bimanual YAM robots, 2.3 times the alternative action representations^{Source 31}31 Zha et al. 2026, EgoLAP: Learning from Egocentric Human Data through Language-Action Reasoning. Lihan Zha, Shresth Grover, Tenny Yin et al., 2026.

Phantom needs no robot data: it inpaints the human arm, overlays a rendered robot and deploys zero-shot, with success up to 92%^{Source 23}23 Lepert et al. 2025, Phantom: Training Robots Without Robots Using Only Human Videos. Marion Lepert, Jiaying Fang, Jeannette Bohg, CoRL 2025. Masquerade robotizes in-the-wild human video and fine-tunes on 50 robot demonstrations per task, beating baselines by 5 to 6 times^{Source 24}24 Lepert et al. 2026, Masquerade: Learning from In-the-wild Human Videos using Data-Editing. Marion Lepert, Jiaying Fang, Jeannette Bohg, ICRA 2026. Figure reports that Index human-video pretraining, all else fixed, raised Helix 2.5's zero-shot success in 30 unseen homes from 9% to 56%^{Source 25}25 Figure AI 2026, Helix 2.5: Zero-Shot 30-Home Generalization. Figure AI, Figure news, 2026, and Generalist's GEN-1 base model pretrains on over half a million hours from wearable devices on humans, with no robot data^{Source 29}29 Generalist Team 2026, GEN-1: Scaling Embodied Foundation Models to Mastery. Generalist Team, Generalist AI blog, 2026.

Scaling the number of embodiments

Across about 1,000 procedurally generated locomotion bodies, adding embodiments generalizes more broadly than adding data on fixed ones^{Source 26}26 Ai et al. 2025, Towards Embodiment Scaling Laws in Robot Locomotion. Bo Ai, Liu Dai, Nico Bohlinger et al., CoRL 2025. Skild reports training one model on 100,000 simulated robots and testing zero-shot on robots excluded from training^{Source 27}27 Skild AI Team 2025, The case for an omni-bodied robot brain. Skild AI Team, Skild AI blog, 2025. DreamZero adapted from AgiBot G1 to a YAM bimanual robot with 55 trajectories, about 30 minutes of play data, though its authors describe the result only qualitatively^{Source 30}30 Ye et al. 2026, World Action Models are Zero-shot Policies. Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng et al., 2026.