Bimanual Robot Manipulation via Multi-Agent In-Context Learning

2026-04-23 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

BiCICLe (Bimanual Coordinated In-Context Learning) is a novel framework enabling Large Language Models (LLMs) to perform few-shot bimanual robot manipulation without fine-tuning. It addresses the challenge of high-dimensional joint action spaces and tight inter-arm coordination by framing bimanual control as a multi-agent leader-follower problem, decoupling the action space into sequential, conditioned single-arm predictions. The framework introduces "Arms' Debate" for iterative refinement and an "LLM-as-Judge" for Best-of-N self-evaluation to select plausible coordinated trajectories. Evaluated on 13 tasks from the TWIN benchmark, BiCICLe achieves up to a 71.1% average success rate, outperforming the best training-free baseline by 6.7 percentage points and surpassing most supervised methods. It also demonstrates strong few-shot generalization on novel tasks not included in the benchmark.

Key takeaway

For research scientists developing bimanual robot control systems, BiCICLe offers a compelling training-free alternative to traditional supervised methods. You should consider adopting its leader-follower decomposition and inference-time refinement strategies to achieve strong performance and out-of-distribution generalization, especially for tasks requiring precise inter-arm coordination. This approach significantly reduces the need for extensive, task-specific datasets and fine-tuning, accelerating deployment to novel scenarios.

Key insights

Decoupling bimanual robot control into leader-follower LLM agents enables effective, training-free in-context learning.

Principles

• Factor bimanual control into sequential single-arm predictions.
• Explicitly condition follower arm on leader's plan.
• Iterative refinement improves inter-arm synchronization.

Method

BiCICLe uses a leader-follower LLM architecture where one arm predicts its trajectory, then the other conditions on it. "Arms' Debate" iteratively refines plans, and "Best-of-N" uses an LLM-as-Judge to select optimal trajectories.

In practice

• Use text-based object positions for robust observations.
• Employ voxel downsampling for accurate point cloud centroids.
• Consider leader-follower for asymmetric manipulation tasks.

Topics

• Bimanual Robot Manipulation
• In-Context Learning
• Multi-Agent LLM Coordination
• BiCICLe Framework
• Leader-Follower Decomposition

Best for: Research Scientist, AI Scientist, Robotics Engineer, Machine Learning Engineer

Related on AIssential

• Bian Que: An Agentic Framework with Flexible Skill Arrangement for Online System Operations — Bian Que orchestrates LLM agents for online system operations by flexibly arranging skills and enabling self-evolution.
• DuoBench: A Reproducible Benchmark for Bimanual Manipulation in Simulation and the Real World — DuoBench offers a reproducible benchmark for bimanual robot manipulation, revealing current policy limitations in coordination and sim-to-real transfer.
• Agentic MPC for Semantic Control System Resynthesis — An agentic MPC framework integrates LLM-based agents to enable context-aware, semantically adaptive control synthesis.
• Factored Diffusion Policies:Compositionally Generalized Robot Control with a Single Score Network — Factored Diffusion Policies use a single diffusion network with additive score decomposition for compositionally generalized robot control.
• AIhub monthly digest: February 2026 – collective decision making, multi-modal learning, and governing the rise of interactive AI — AIhub's February 2026 digest highlights progress in multi-agent systems, robot learning, neurosymbolic AI, and governance.
• InterPrior: Scaling Generative Control for Physics-Based Human-Object Interactions — InterPrior scales generative control for human-object interactions using imitation learning, data augmentation, and reinforcement learning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.