Reproduce “A Dual Process VLA: Efficient Robotic Manipulation Leveraging VLM” on OpenResearch

About this paper

ByungOk Han, Jaehong Kim, Jinhyeok Jang

The Dual Process VLA (DP-VLA) framework introduces a hierarchical approach to robotic manipulation, balancing the reasoning of large vision-language models with real-time control by separating high-level planning from low-level execution. This architecture, implemented using OpenVLA as the L-Sys2 and BC-Transformer as the S-Sys1, achieved higher task success rates and faster inference speeds of over 10 Hz in the RoboCasa simulation environment.

Key insights

• Latent features extracted from the high-level reasoning model (L-Sys2) remain valid for multiple timesteps, allowing the low-level control model (S-Sys1) to operate continuously and efficiently without constant updates from the computationally intensive L-Sys2.
• Separating high-level reasoning from real-time control enables a VLA system to balance sophisticated reasoning capabilities with the need for efficient, real-time robot operation.
• Latent features from the decoding stage of the L-Sys2 (OpenVLA) and features from its pre-trained state proved more effective for transferring knowledge to the S-Sys1 than features from a fine-tuned L-Sys2.

Summary by alphaXiv.