The 2026 IEEE International Conference on Robotics & Automation (ICRA) was held in Vienna from June 1 to 5, bringing together robotics experts. IEEE has made keynote and plenary talk recordings available to view, along with panel discussions and tutorials. Topics covered included agentic coding, sustainable robotics, aerial robots, and human-robot interaction. Various panels addressed funding, sustainability, AI for automation, and global visions for robotics.
Maven Robotics is leading an investment in Series A, focusing on building general-purpose robotics for enterprise customers. The core challenge in robotics is achieving large-scale, reliable work that satisfies stringent enterprise requirements, which demands multidisciplinary expertise across the entire technology stack. Maven addresses this by adopting a "deep, then wide" strategy, focusing on building a general platform capable of mastering specific tasks to deliver immediate customer value. The company's team combines deep experience from major technology and aerospace companies, emphasizing industrialization in their approach. Maven aims to deliver solutions that meet demanding enterprise KPIs across complex logistics and handling tasks.
Bi-MoDe is a modifier-conditioned decoding framework for bilateral control-based imitation learning. It introduces a constrained latent into the Transformer action decoder via adaLN-Zero to modulate execution speed and contact intensity. The method directly influences action-chunk generation based on behavioral directives. Evaluation on a whiteboard wiping task showed Bi-MoDe improved physical directive following compared to a baseline while maintaining temporal control. The interaction between decoder conditioning and latent-space composition was found to be important for accurate physical directive following.
MR-GLi is a Mixed Reality-Based interface designed for underwater robot arm teleoperation using bilateral control. It spatially registers a reaction torque indicator and wrist-camera image directly onto the robot gripper. Experiments demonstrated that this interface provided gripper-linked access to visual feedback while maintaining comparable torque-regulation performance to a 2D monitor. Subjective evaluations also showed reduced cognitive burden for users shifting attention between the workspace and visual feedback. These findings confirm the feasibility of gripper-linked MR overlays for underwater teleoperation.
World-Action Models (WAMs) combine future world prediction with executable action generation for robots operating in complex environments. This survey reviews WAMs relative to other methods like world models and reinforcement learning. It organizes existing methods across representations, transition modeling, action interfaces, and training strategies. The work covers applications in manipulation, navigation, and autonomous driving, reviewing relevant datasets and challenges. Key challenges discussed include action alignment, world-action factorization, and efficient inference for reliable embodied intelligence.
AssemblyGrid v1 is a reproducible benchmark for multi-robot production involving temporary coalitions, local information, and geometric constraints. It incorporates explicit process progression, decentralized observations, material transfer, and productive concurrency into a single task-level formulation. The benchmark includes Flow, Coalition, and Concurrency workload families with multiple scenario levels. It allows evaluation of learning-based and non-learning methods by defining success based on task achievement rather than specific learning rewards. Experiments show that decentralized policies can effectively learn production behavior from local observations.
This work introduces a quantum representation and circuit architecture for rigid-body transformations, specifically applied to the Denavit-Hartenberg kinematics of serial open-chain robot manipulators. Translational contributions are encoded via a classical metric and a single-qubit Bloch vector, while rotations are captured by parameterized frame direction propagation. A selector register prepares probabilities, and readout qubits encode the normalized weighted sum and terminal-frame axes, respectively. Ideal expectation values reproduce classical kinematic quantities, and simulations show negligible reconstruction errors.
A self-contained, modular continuum robotic platform is presented that incorporates mechanical reconfigurability and in-situ self-pose estimation. The robot uses interchangeable continuum joints with precomputed stiffness for rapid assembly and shape programming. Proprioceptive sensing is achieved via magnetic sensors and a modular, joint-specific learning-based framework for pose estimation. Experimental validation demonstrated the system's ability to self-sense and adapt without external tracking. This work enables more adaptable and deployable continuum robots.
ConGraspXL extends GraspXL by introducing controllable constraint-conditioned grasping motion synthesis. It addresses the limitation of prior work by accommodating diverse, task-driven constraints and their combinations. The method employs a hierarchical constraint formulation and a masked residual interface for flexible constraint composition. Precision is enhanced through dynamic hand centers and feed-forward wrist guidance. This results in a plug-and-play low-level grasp controller useful for tasks like whole-body grasp completion.
Auto-HSI generates personalized human-swarm interaction (HSI) interfaces on demand using Large Language Models (LLMs) for automatic code generation. This method allows untrained operators to use natural language and gestures to instruct robots on desired collective behavior. The generated code creates personalized state machines enabling centralized control via one- and two-handed gestures for motion and shape manipulation. Experiments demonstrated successful gesture tracking and code generation, followed by live control of 50 simulated robots in a physics-based simulator. The system was validated through tasks involving goal scoring, maze traversal with shape deformation, and group splitting under nominal and noisy conditions.