New TAD method distills VLA policies using only teacher trajectories

Researchers have developed Trajectory-only Adversarial Distillation (TAD), a novel method for training vision-language-action (VLA) policies. TAD enables policy distillation using only teacher trajectory features, bypassing the need for explicit teacher action scores. This approach significantly improves the performance of models like OpenVLA-OFT on complex tasks, achieving a substantial increase in success rates on the ManiSkill benchmark. TAD demonstrates a more efficient use of interaction and optimization resources compared to traditional methods. AI

IMPACT Enables more efficient training of complex AI policies without requiring explicit teacher action scores.

RANK_REASON Academic paper detailing a new method for training AI policies. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• ManiSkill
• OpenVLA-OFT
• PutOnPlateInScene25Main
• Trajectory-only Adversarial Distillation

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →