OASIS: From Simulation Data Collection to Real-World Humanoid Loco-Manipulation
AuthorsZehao Yu, Jiakun Zheng, Weiji Xie, Jiyuan Shi, Chenyun Zhang, Chenjia Bai, Xuelong Li
Resources
OASIS shows that carefully built simulated data can sometimes beat real teleoperation for humanoid robot loco-manipulation on the physical robot.
Key results
OASIS vs real-robot teleoperation for kneel-and-wipe-under-table
Minutes for 50 successful place-cup-in-box trajectories with OASIS
Minutes for 50 successful kneel-and-wipe-under-table trajectories with OASIS
Full domain-randomized rendering pipeline with 20 environments per trajectory
Ablation without any visual randomization
What the paper found
OASIS, from China Telecom’s TeleAI and collaborators at Fudan University, East China University of Science and Technology, and Shanghai Jiao Tong University, is a simulation-first pipeline for humanoid loco-manipulation that replaces real-robot teleoperation with scalable simulation data. It reconstructs physics-ready objects from single-view real images using Hunyuan3D and Qwen3-VL, then lets operators teleoperate a simulated Unitree G1 through PICO 4 Ultra VR, recording only state sequences for later offline photorealistic replay with texture, lighting, and camera-extrinsics randomization. The policy stack uses a hierarchical design: a Flow Matching high-level planner conditioned on frozen CLIP text features, DINOv2 multi-view images, and 2-frame proprioceptive history predicts 32-frame reference motion chunks, which a low-level controller converts into 43-DoF whole-body joint angles. In real-robot deployment, OASIS cuts data-collection time to 15.2, 19.1, 25.2, and 28.4 minutes for 50 successful trajectories across four tasks, versus 17.5, 26.8, 40.2, and 44.8 minutes for real teleoperation, reaching up to 1.84× speedup. Zero-shot transfer is especially sensitive to visual augmentation: removing all randomization drops average success to 0.05, while the full pipeline reaches 0.83 after rendering each trajectory in 20 randomized environments. On the Unitree G1, simulation-only training matches or exceeds real-data training on tasks such as place-cup-in-box, lift-basket-and-place-cup, wipe-monitor, and kneel-and-wipe-under-table, showing that high-fidelity simulation can substitute for expensive real-world collection and improve robustness through broader visual coverage.
Original abstract
Recent progress in robot manipulation has been largely driven by learning from large-scale demonstrations. For humanoid robot loco-manipulation tasks, however, existing data sources force an unsatisfying tradeoff between trajectory quality and scalability. Real-world teleoperation provides the highest-quality trajectories but requires dedicated physical space and time-consuming scene resets. Simulation offers an alternative way out of this dilemma: it can produce clean, embodiment-aligned data at scale without any physical hardware. In this paper, we propose OASIS, a simulation-data-driven framework for humanoid loco-manipulation. OASIS automatically reconstructs realistic object assets from real-world images using a 3D generative model. Based on these assets, trajectories are first collected through teleoperation in simulation, and then augmented under diverse domain randomizations in a post-processing stage. With the resulting simulation data, we further design a hierarchical visuomotor policy for humanoid loco-manipulation. Extensive experiments on the real humanoid robot show that, under zero-shot deployment, the policy trained on our simulation data achieves higher success rates on most tasks than that trained on real-robot teleoperation data, owing largely to the broad lighting and environmental variations covered by our simulation rendering, which real-robot data fails to capture. The project page is available at https://oasis-humanoid.github.io/.
Read the original paperMore in Robotics
Browse all 50 papers →JAMB: Joint Action-Motion Diffusion for Bimanual Manipulation
Chuyang Xiao, Peilin Meng, David Held
JAMB helps two robot arms coordinate by jointly imagining their future movements and the changing 3D scene before acting.
Rolling-WAM: World Action Models with Rolling Imagination
Yinghua Zhou, Junjie Ye, Yiqi Zhao, Hao Dong, Celina Shiyu Wang, Ruohai Ge, Tingyi Yang, Basile Van Hoorick, Gaurav Sukhatme, Vitor Guizilini, Yue Wang
Rolling-WAM keeps future robot actions partially imagined and refined over time, making world-model-based manipulation replan 4.5 times faster.
Training-free Behavior Cloning
Maximilian Adang, Timothy Chen, Lars Osterberg, Aiden Swann, Mac Schwager
A fast, training-free robot controller reuses and corrects demonstration trajectories to deliver traceable behavior at real-time speeds.