Researchers have developed WAM-Policy, a novel approach to enable quadrupedal robots to perform loco-manipulation tasks guided by language instructions. This method adapts existing World Action Models (WAMs), which are typically trained for robot arms, to quadrupedal robots. The process involves two stages: initial adaptation and acceleration using on-policy self-distillation, followed by task-specific improvement and merging through multi-teacher on-policy distillation. This technique significantly enhances policy performance and reduces inference costs, as demonstrated on the QLM-Bench and a real Unitree Go2 Air robot. AI
IMPACT Enables quadrupedal robots to perform complex tasks guided by natural language, potentially broadening their applications in unstructured environments.
RANK_REASON The cluster describes a research paper detailing a new method for robot control. [lever_c_demoted from research: ic=1 ai=1.0]
- Cross-Morphology Policy Adaptation and Acceleration
- Multi-Teacher On-Policy Distillation
- On-Policy Distillation
- On-policy self-distillation
- QLM-Bench
- Quadrupedal Loco-Manipulation
- reinforcement learning
- Task-Specific Improvement and Merging
- Unitree Go2 Air
- WAM-Policy
- World Action Models
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →