Adapt as You Say: Online Interactive Bimanual Skill Adaptation via Human Language Feedback

Generalizable bimanual robotic manipulation requires a reusable task prior that can persist across increasingly diverse tasks, objects, scenes, embodiments, and execution conditions, thus avoiding the prohibitive cost of large-scale teleoperated demonstrations and policy retraining. In this work, we present VLBiMan++, an extended framework that expands the generalization boundary of vision-language anchored one-shot bimanual manipulation. Starting from a single human demonstration, VLBiMan++ performs task-aware decomposition to identify reusable and adaptable skill components, and employs vision-language grounded geometric adaptation to transfer these skills to novel configurations without retraining. Building on this foundation, we systematically extend generalization along five dimensions: task generalization through diverse and long-horizon skill compositions; object generalization across unseen categories, varying geometries, and more complex articulated or deformable objects; scene generalization under clutter, occlusion, and dynamic interference; embodiment generalization across heterogeneous dual-arm robotic platforms; and deployment generalization through prolonged closed-loop execution under repeated external perturbations. To support this broader scope, we further introduce object-state-aware adaptation and lightweight trajectory optimization mechanisms that accommodate changes beyond simple rigid 6-DoF pose variations while preserving reliable bimanual coordination. Extensive real-world experiments demonstrate that VLBiMan++ maintains strong task success and adaptation capability across these increasingly challenging settings. Overall, VLBiMan++ advances one-shot bimanual manipulation from demonstrating isolated transferability toward a more systematic and scalable framework for generalization across tasks, objects, scenes, embodiments, and long-term deployment conditions.

Robust manipulation under uncertain object dynamics and geometry remains a core challenge in real-world robotics. Policies trained under fixed assumptions often degrade when object mass, friction, scale, or shape vary. Embedding-based adaptation methods such as Rapid Motor Adaptation (RMA) address this by conditioning actions on latent representations of environment dynamics inferred from interaction history. We investigate whether this paradigm scales from single arm settings to coordinated bimanual manipulation, where higher dimensional control, coupled contacts, and shared object forces introduce additional complexity. We evaluate bimanual RMA on three contact-rich ManiSkill tasks with randomized YCB object properties. Bimanual RMA, using observation-action history and dual wrist depth images, retains privileged policy performance on bimanual picking and lifting, while an otherwise matched stationary PPO policy without latent conditioning achieves low success; all methods remain limited on long-horizon stacking. We further demonstrate direct sim-to-real transfer on a dual Franka lifting task without real-world finetuning. Together, these results suggest that latent conditioning provides useful structure for scaling robust control to coordinated bimanual systems, while longer horizon tasks require additional temporal structure.

Large Language Models (LLMs) have emerged as powerful reasoning engines for embodied control. In particular, In-Context Learning (ICL) enables off-the-shelf, text-only LLMs to predict robot actions without any task-specific training while preserving their generalization capabilities. Applying ICL to bimanual manipulation remains challenging as the high-dimensional joint action space and tight inter-arm coordination constraints rapidly overwhelm standard context windows. To address this, we introduce BiCICLe (Bimanual Coordinated In-Context Learning), the first framework that enables standard LLMs to perform few-shot bimanual manipulation without fine-tuning. BiCICLe frames bimanual control as a multi-agent leader-follower problem, decoupling the action space into sequential, conditioned single-arm predictions. Evaluated on 13 tasks from the TWIN benchmark, BiCICLe achieves 70.5% average success rate, outperforming the best training-free baseline by 6.1 percentage points and surpassing most supervised methods. We also demonstrate superior real-world performance on 3 tasks without hardware-specific retraining. The project page is available at https://alesspalma.github.io/bicicle