Researchers have developed a Quality-Controllable Policy (QCP) for robots, enabling dynamic control over execution attributes like speed and smoothness at inference time. To expand the range of controllable qualities beyond initial demonstrations, QCP employs an iterative hindsight relabeling loop. This process involves collecting online rollouts, adding successful trajectories to the dataset, relabeling them based on measured quality, and fine-tuning the policy. Experiments on Robomimic and RoboEval show that QCP enhances both task execution quality and success rates, while real-world tests demonstrate practical gains in physical deployments, such as increased throughput and improved stability. AI
IMPACT Enables robots to adapt execution qualities like speed and smoothness in real-time, improving efficiency and success rates in physical tasks.
RANK_REASON Academic paper detailing a new method for robot policy control. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
