aiqbench analysis
This research introduces MoE-ACT, a parameter-efficient framework that leverages sparse Mixture-of-Experts layers to solve task interference in multi-task bimanual robotic manipulation. By dynamically routing visual tokens and utilizing task-conditioned modulation, the model achieves superior performance on the RoboTwin 2.0 benchmark while using significantly fewer activated parameters than standard foundation models. The results demonstrate a viable path toward deploying scalable, unified visuomotor policies for complex real-world robotic applications.
MoE-ACT: Scaling Multi-Task Bimanual Manipulation with Sparse Task-Conditioned Mixture-of-Experts Transformers
arXiv:2603.15265v2 Announce Type: replace Abstract: Developing a unified policy for multi-task robotic manipulation remains challenging due to policy degradation from task interference and negative transfer. In this work, we propose Mixture-of-Experts-Enhanced Action Chunking Transformer (MoE-ACT), a parameter-efficient multi-task visuomotor framework tailored for bimanual manipulation. MoE-ACT incorporates sparse MoE layers into the ACT encoder, dynamically routing image tokens to selected experts based on task context, visual observations, and proprioceptive states. Furthermore, the framework incorporates task-conditioned Feature-wise Linear Modulation (FiLM) in the action decoder alongside multi-scale cross-attention, ensuring precise task grounding and capturing fine-grained spatial cues. Extensive evaluations on the RoboTwin 2.0 benchmark across 16 challenging bimanual tasks demonstrate that MoE-ACT achieves an average success rate of 62.0%, outperforming standard ACT by 17.4 percentage points. Crucially, with only 195M activated parameters, MoE-ACT surpasses the 16-fold larger foundation model $\pi$0 (3.24B) by 5.1 percentage points, exhibiting substantial gains in parameter efficiency and deployment feasibility. Real-robot dual-arm experiments consistently confirm its superior multi-task execution capabilities. Our open-source project page can be found at https://j3k7.github.io/MoE-ACT/.
Abstract:Developing a unified policy for multi-task robotic manipulation remains challenging due to policy degradation from task interference and negative transfer. In this work, we propose Mixture-of-Experts-Enhanced Action Chunking Transformer (MoE-ACT), a parameter-efficient multi-task visuomotor framework tailored for bimanual manipulation. MoE-ACT incorporates sparse MoE layers into the ACT encoder, dynamically routing image tokens to selected experts based on task context, visual observations, and proprioceptive states. Furthermore, the framework incorporates task-conditioned Feature-wise Linear Modulation (FiLM) in the action decoder alongside multi-scale cross-attention, ensuring precise task grounding and capturing fine-grained spatial cues. Extensive evaluations on the RoboTwin 2.0 benchmark across 16 challenging bimanual tasks demonstrate that MoE-ACT achieves an average success rate of 62.0%, outperforming standard ACT by 17.4 percentage points. Crucially, with only 195M activated parameters, MoE-ACT surpasses the 16-fold larger foundation model $\pi$0 (3.24B) by 5.1 percentage points, exhibiting substantial gains in parameter efficiency and deployment feasibility. Real-robot dual-arm experiments consistently confirm its superior multi-task execution capabilities. Our open-source project page can be found at this https URL.
Robotics (cs.RO)
arXiv:2603.15265 [cs.RO]
(or arXiv:2603.15265v2 [cs.RO] for this version)
https://doi.org/10.48550/arXiv.2603.15265 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 258 words
Related coverage
STAR: Sparse Tactile Representation Learning in Vision-Tactile-Language-Action Models for Dexterous Manipulation
arXiv · Sep 14, 2026
LiMA: Bridging Long-term Imagination to Real-time Dexterous Manipulation via Asynchronous Diffusion
arXiv · Sep 24, 2026
Gemini 4 Argon for Robotics
Google DeepMind · Oct 13, 2026
Scaling Bimanual Household Manipulation from 1,500 hours of Demonstrations to On-Policy Corrections
arXiv · Sep 4, 2026