SUGARv2: Learning Humanoid Mobile Manipulation from Synthesized Interaction References

Tianshu Wu and other authors released a paper proposing the SUGARv2 framework for learning diverse visual humanoid mobile manipulation skills. The method adapts reusable task templates to different spatial configurations, synthesizes low-cost references of coordinated robot-object motion, and uses reference-based imitation as a shared learning objective across tasks: it first trains a privileged teacher to execute these references, then trains a visual policy via teacher distillation and reinforcement learning, while using diverse object instances and visual domain randomization to support visual generalization and sim-to-real transfer. The authors report that across five humanoid mobile manipulation tasks, SUGARv2 achieves an average success rate of 91.0% in simulation; after deployment on a real humanoid robot, whole-body manipulation also achieved sim-to-real transfer. The above success rate and transfer results are all results reported by the paper's authors.

AI 根据报道生成 · 2 days ago更新