V-PTR
V-PTR
Robotic Offline RL from Internet Videos
Robotic Offline RL from Internet Videos
via Temporal-Difference Learning
via Temporal-Difference Learning
Video
Video
V-PTR.mp4About the System
About the System
Pre-trains on internet video data with value-based RL to get functional representations to understand possible future outcomes
Pre-trains on diverse robot data to understand the actions that lead to future outcomes
Fine-tunes on task-specific robot demos to understand how to complete specific tasks
Results + Rollouts
V-PTR (ours)
final_croissant_ego4d_left_rear_shift_2_success.mp4PTR
final_croissant_scratch_left_front_shift_2_failure.mp4Masked Visual Pre-training
final_croissant_bc_mae_left_front_shift_1_failure.mp4VIP
final_croissant_rl_baselines_vip_left_rear_under_failure.mp4R3M
final_croissant_bc_r3m_right_front_shift_2_failure.mp4Full Paper + Supplementary
Full Paper + Supplementary
VPTR_NeurIPS_WS_9p.pdf