Reimagine Robotics is developing technology that allows robots to learn on the job through human instruction, bypassing specialist programmers for task changes. The startup claims its system enables workers to show the robot what to do, watch it attempt the task, and correct it immediately. This approach relies on people remaining central to the process, with humans identifying bottlenecks, teaching robots, and correcting them until they become useful. Reimagine Robotics has already deployed its robots in advanced manufacturing and electronics disassembly facilities, reducing prototype testing time from a day to 10 minutes.
Google DeepMind's Gemini Robotics 2 enables full-body control of humanoid robots like Apptronik's Apollo 2, allowing them to reason through every movement and unlock a broad range of tasks. The model can manage complex tasks that require multiple steps over an extended period, self-correct if a step fails, and generalize to novel situations and goals. Gemini Robotics ER 2 is now available on Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, while Gemini Robotics On-Device 2 targets applications with low connectivity. The model can adapt to new robot embodiments with just a few hours of adaptation time.
A new package called ros2_realtime_support provides real-time functionality for ROS 2 without modifying existing packages. The package is designed to be used with rclcpp, and users can enable real-time support by adding a configuration file, changing the main routine in source files, and updating certain classes. However, the development of this package is still ongoing, with future work planned to improve mutexes and condition variables on RTOS.
Humanoid and quadrupedal robots have the potential to revolutionize various aspects of society, but current capabilities along hardware, locomotion, autonomy, data, and applications need assessment. Recent advances in legged robots include improved balance control, advanced materials, and machine learning algorithms. However, key challenges must be overcome, such as increased energy efficiency, stability, and adaptability. The authors provide an outlook on the future of legged robots, exploring their potential for scientific discovery, economic benefits, policy implications, and broader societal effects.
ST-WAM, a semantic-temporal World Action Model, improves action robustness under visual distribution shifts by using DINOv3 as a shared semantic representation for future prediction and history retrieval. It jointly predicts future VAE latents and DINO features, while retrieving task-relevant evidence from recent DINO history. ST-WAM achieves 98.7% on LIBERO and 92.8% on RoboTwin 2.0, improving zero-shot performance by 21.3 percentage points compared to Fast-WAM.
D-VLC, a decentralized vision-language collaboration framework, enables heterogeneous embodied multi-robot systems to collaborate in unknown environments. It combines decentralized asynchronous reasoning, lightweight information sharing, capability aware collaboration, and a unified action interface, allowing general purpose VLMs to generate robot-specific actions without task or robot-specific training. Experiments show success rates above 70% and completion time reductions of up to 55.8% relative to the geometric greedy baseline.
Auto-JEPA, an action-oriented latent world model, predicts continuous future driving intent through joint-embedding prediction. It learns to focus on scene features that affect future ego action by keeping the visual encoder frozen and requiring no explicit perception annotations. Auto-JEPA achieves 91.3 PDMS on NAVSIM v1 and 89.1 EPDMS on NAVSIM v2, outperforming dense future-world modeling without focusing on planning-relevant visual features.
Researchers at Yiting Mo and co-authors have developed a variable-stiffness tactile interface (VSTaI) based on 3D-printed structured fabrics. The VSTaI uses vacuum-induced jamming to increase stiffness, achieving an effective modulus from sub-megapascal to megapascal levels. Four fabric patterns with different geometric parameters were fabricated and evaluated using force-indentation tests under atmospheric and vacuum conditions, showing a significant increase in stiffness when vacuum jammed. This technology supports structured-fabric jamming as a practical approach for shape-conformable, tunable-stiffness displays aimed at physical examination training.
ActFovea, a plug-and-play safeguarding framework, detects and mitigates runtime disturbances in vision-language-action (VLA) policies. It uses robot kinematics, proprioceptive states, and recent actions to construct action-conditioned foveated regions that retain contact-relevant areas and predicted motion corridors while suppressing task-irrelevant visual content. ActFovea increases success under localized visual overlays from 49.3% to 90.3%, closing 93.7% of the gap to clean performance, while preserving clean-task performance.
CLIFT: Turning Gemini Robotics On-Device into Humanoid Specialists via Non-Invasive Closed-Loop Iterative Fine-Tuning introduces CLIFT, a method that enables closed-loop policy improvement without accessing model weights or gradients. The approach turns deployment-time reward feedback into API-compatible supervised data, achieving near-perfect success on agile, contact-rich tasks after two flywheel cycles.