VQActFlow: Vector-Quantized Action Mode Steering for Multi-Task Robot Manipulation

2026-06-19

AI Summary

This work addresses the challenge of learning multi-task robotic policies from multimodal demonstrations, where a single policy network must switch between distinct action modes—a process prone to task failure or infeasible actions. To overcome this, the authors propose the first integration of vector quantization with diffusion flow matching, encoding continuous actions into discrete codebook tokens and generating sequences of action codes. By explicitly modeling preferences over action modes, the approach incorporates classifier-free guidance and a codebook critic to jointly modulate semantic alignment with language instructions and scene feasibility. Evaluated on the LIBERO simulation benchmark, full-body manipulation with the Unitree G1 humanoid robot, and contact-rich bimanual tasks using ALOHA, the method consistently outperforms both continuous and discrete baselines, significantly improving mode selection accuracy and task generalization.

Abstract

Multi-task robot manipulation policies are challenging to learn from demonstration because traditionally a single network must select among qualitatively different action modes from a multimodal demonstration distribution, conditioned on language and visual context. A wrong mode selection means executing the wrong task or an action infeasible in the scene. Tokenizing continuous actions into a learned discrete codebook separates these modes at the representation level, offering structural advantages for multi-task learning. We propose VQActFlow, a multi-task manipulation policy that tokenizes action chunks and generates code sequences via Variational Flow Matching. VQActFlow maintains an explicit preference over action modes throughout generation. Inference-time guidance acts on this preference to steer mode commitment. We instantiate this with classifier-free guidance over language conditioning, which steers the policy toward the instructed action mode, and a learned codebook critic that supplies a complementary feasibility signal. We evaluate VQActFlow on three platforms: the LIBERO simulation benchmarks, a Unitree G1 humanoid performing whole-body pick-and-place, and an ALOHA-style bimanual platform performing contact-rich tasks. Across these benchmarks, VQActFlow outperforms both continuous and discrete baselines.

Problem

• multi-task robot manipulation
• action mode selection
• multimodal demonstration
• discrete action representation
• language-conditioned control

Innovation

• Vector Quantization
• Action Mode Steering
• Variational Flow Matching
• Multi-Task Manipulation
• Discrete Action Tokenization