The authors developed DRILL, a two-level training framework for vision-language-action policies, which translate visual observations and language instructions into actions.
Training world models for policy improvement beats optimizing pixel accuracy
Across five manipulation simulators and two robot-policy backbones, the authors report higher task success from models trained to improve learning updates rather than next-frame predictions.
Chinese Tech
Zhaolu Kang · Tailong Luo · Lei Wei · Guangyuan Dong · Sihan Liu · Yingjie He · +6 more
Hunyuan, Tencent · Peking University
Research Digest··2 min read
Kang et al.
§