China is prioritizing embodied AI and humanoid robotics to achieve industrial and geopolitical dominance, leveraging its manufacturing base and state-orchestrated Five-Year Plans. These plans drove the robotics sector, leading to significant increases in domestic industrial robot market share and global installation quantities by 2025. A critical bottleneck remains in precision hardware and semiconductor self-sufficiency, as reliance on foreign components hinders full autonomy. The 15th Five-Year Plan is shifting focus to embodied AI, heavily subsidizing humanoid development to drive physical AI and data aggregation. This strategy involves a high-risk gamble on humanoids, balancing rapid development with the risk of market overcapacity and structural instability.
NVIDIA Isaac ROS 5.0 introduces GPU-accelerated packages and agentic workflows to accelerate the development and deployment of robotics applications. It integrates NVIDIA accelerated computing, physical AI models, and production-ready libraries into the open-source ROS framework. This release provides agent-ready skills, such as FoundationPose for faster object perception and agent-ready pick and place workflows. The update supports the latest ROS platforms and leverages CUDA for efficient GPU acceleration across different hardware. It enables the combination of an open robotics ecosystem with accelerated computing to build capable physical AI systems.
ROS 2 nodes can now exchange GPU-resident payloads via zero-copy transport using the `rosidl::Buffer` and the NVIDIA CUDA buffer backend. This backend allows for efficient data movement between nodes without serialization or host copies when runtime conditions permit. An AI coding agent, using the `migrate-node-to-rosidl-buffer` skill, automates the process of auditing, refactoring, and verifying this migration in CUDA-accelerated nodes. The process demonstrated on a Depth Anything 3 node involves updating data handling to use CUDA buffers for direct inference results. This enables faster, more efficient data pipelines in robotics applications on platforms like NVIDIA Jetson AGX Thor.
Barbara Mazzolai seeks to establish a new field of research called "sustainability robotics" by integrating environmental responsibility into robot design. She draws inspiration from nature, using biological models like octopuses and plant roots to develop innovative robotic systems. Her vision emphasizes reducing technological footprints and ensuring robots are symbiotic with both humanity and nature. Key to her approach is incorporating a life cycle perspective so robots can be reused, recycled, or biodegraded. Mazzolai aims to inspire future roboticists to develop technology that actively helps mitigate environmental damage.
The study proposes a Deep Reinforcement Learning (DRL) strategy to enable mobile human-companion robots to dynamically adjust their tracking positions in real-world environments. This method addresses the limitations of traditional methods that restrict robots to fixed relative positions. A human-robot companion controller is developed by integrating Model Predictive Path Integral (MPPI) control with Control Barrier Functions (CBF) for safe and accurate following. Experiments demonstrated that the proposed approach improved tracking accuracy and success rates by 24% and 47%, respectively, while enhancing human comfort. The robot successfully accompanied a person walking at speeds up to 1.7 m/s while respecting intimate space and obstacles.
This paper formulates a separated-section constitutive law for cosserat rod models of trimmed helicoid soft arms. It addresses the inaccuracy of standard models for these arms by accounting for separated load-bearing helix domains connected by sparse fused crossings. The model captures compliance from relative motion between domains, resulting in a strongly anisotropic effective sectional stiffness. This law is integrated into a geometrically exact dynamic Cosserat model suitable for rapid planning and control in soft robots. Experimental results show small position errors across measured configurations.
The paper proposes a reinforcement learning (RL) approach to automatically generate multimodal policies for robot assistants in human-robot collaboration. The system focuses on a scenario where a robot assists a user in object location by managing language and physical action signals. The agent is trained using a simulator incorporating human data and a simple reward function. Human study results show the system demonstrates high usability and effective task completion in real-world settings. This method offers a scalable alternative for designing interaction managers in multimodal collaborations.
Coordinated bimanual manipulation is challenging because arm motions affect the shared 3D scene. JAMB proposes a diffusion policy that jointly denoises bimanual actions and future 3D point tracks. It achieves this by evolving action and track hypotheses together within a shared Transformer. The method grounds multimodal representations in a shared spatiotemporal coordinate system for geometry-aware interaction. JAMB significantly outperforms baselines across simulation and real-world bimanual manipulation tasks.
GINIO introduces a geometric SO(3)-equivariant interface for neural inertial odometry to handle arbitrary IMU frame rotations. The framework ensures that motion measurements and uncertainties transform consistently according to tensorial laws. It includes Last-Frame Alignment (LFA) for efficient sensor-frame learning and a connected estimator to track nuisance parameters. Experiments show GINIO achieves superior performance in various benchmarks, significantly reducing absolute time error and covariance-equivariance error compared to other methods.
The research focuses on human-robot collaborative transport, where a robot assists a human in moving objects. The key insight is that obstacle-constrained transport requires integrating predictions of human behavior with compliant robot control. The PROACT framework uses a transformer architecture trained on human transport demonstrations to predict future object motion. This approach resulted in significant reductions in interaction work and completion time compared to baseline methods. PROACT achieved a 59.2\% reduction in interaction work and a 12.9\% reduction in completion time in experiments.
HOTICE is a whole-body humanoid learning framework designed for transporting objects in cluttered environments. It introduces Humanoid-Object Decoupled Potential Fields to guide collision-avoidance for both the robot and the payload. A dual-agent reinforcement learning architecture is used to manage the large action space by decoupling upper- and lower-body control while maintaining whole-body coordination. A specialist-to-generalist distillation strategy is employed to train a policy that generalizes across diverse cluttered scenes. Evaluations in simulation and on a real humanoid demonstrate effective and robust object transportation in complex settings.
Shared control frameworks often over-help due to overconfidence in inferred human intent. This work proposes a capability-aware framework to mitigate this by combining semantic-intent confidence from a Vision-Language Model (VLM) with Vision-Language-Action (VLA) capability confidence. The VLA capability confidence is estimated online from the dispersion and instability of stochastic action trajectories. A nonlinear arbitration policy merges these two confidence measures using a sigmoid mapping to adjust robot authority. Evaluation showed this method achieved the highest task success rate (92%) and improved control friendliness compared to other shared-control methods.