DORA: Dexterous Offline Reinforcement Learning for Vision-Language-Action Models with Integrated Token-Action Critic

1 SJTU 2 Shanghai AI Laboratory 3 USTC
* Equal contribution ** Corresponding author

DORA improves dexterous VLA policies with token-grounded value learning, moving beyond imitation-only behavior and visually brittle external critics.

Abstract

While Vision-Language-Action (VLA) models excel in dexterous manipulation, successful task execution often exhibits multimodality, making reinforcement learning (RL) post-training essential for models to truly understand underlying action intents. However, applying RL is bottlenecked by Visual Value Hacking under severe visual occlusions, alongside credit assignment challenges in generative high-dimensional action spaces. To address this, we propose DORA, an offline RL framework for dexterous VLA models. It leverages a Consistency Policy to drastically shorten the action generation path and introduces a chunk-level Integrated Token-Action Critic. Real-robot experiments show DORA boosts the overall success rate from 40.8% to 60.0% over an imitation-only baseline, with a 25.0 percentage-point improvement on unseen objects and camera occlusion.

DORA Pipeline

Real-world Experiments

To complement the quantitative evaluation, we provide representative real-robot rollout sequences. Each row shows temporally ordered frames from one execution episode, covering successful task executions, object-unseen generalization, and a typical failure case.

(a) Pick and Place

Pick and place rollout frame 1 Pick and place rollout frame 2 Pick and place rollout frame 3 Pick and place rollout frame 4 Pick and place rollout frame 5 Pick and place rollout frame 6 Pick and place rollout frame 7 Pick and place rollout frame 8 Pick and place rollout frame 9 Pick and place rollout frame 10

A successful Pick and Place rollout, illustrating object approach, grasp establishment, transport, and placement.

(b) Pick Various Toys

Pick various toys rollout frame 1 Pick various toys rollout frame 2 Pick various toys rollout frame 3 Pick various toys rollout frame 4 Pick various toys rollout frame 5 Pick various toys rollout frame 6 Pick various toys rollout frame 7 Pick various toys rollout frame 8 Pick various toys rollout frame 9 Pick various toys rollout frame 10

A successful Pick Various Toys rollout, showing adaptation to irregular toy geometry.

(c) Flip Open the Box

Flip open the box rollout frame 1 Flip open the box rollout frame 2 Flip open the box rollout frame 3 Flip open the box rollout frame 4 Flip open the box rollout frame 5 Flip open the box rollout frame 6 Flip open the box rollout frame 7 Flip open the box rollout frame 8 Flip open the box rollout frame 9 Flip open the box rollout frame 10

A Flip Open the Box rollout, demonstrating articulated contact-rich manipulation under strong hand-object occlusion.

(d) Pick Various Toys (Object-Unseen)

Pick various toys object-unseen rollout frame 1 Pick various toys object-unseen rollout frame 2 Pick various toys object-unseen rollout frame 3 Pick various toys object-unseen rollout frame 4 Pick various toys object-unseen rollout frame 5 Pick various toys object-unseen rollout frame 6 Pick various toys object-unseen rollout frame 7 Pick various toys object-unseen rollout frame 8 Pick various toys object-unseen rollout frame 9 Pick various toys object-unseen rollout frame 10

A Pick Various Toys object-unseen rollout, illustrating evaluation on a toy object instance not observed during training.

These qualitative results demonstrate the policy's capability to dynamically adjust topological grasps in contact-rich real-world manipulation, while also revealing current physical limitations under extreme occlusion or under-actuated grasp maintenance.

Comparison

...

We compare the representation structure between an external critic and DORA. The performance gap directly substantiates the necessity of the Integrated Token-Action Critic: relying on raw image inputs and an independent encoder makes the external critic susceptible to visual value deception, while DORA assigns credit on a grounded physical semantic manifold by coupling chunk-level value estimation with VLM cognition tokens.

BibTeX

Coming Soon