ros2_control 6.12.0 removes deprecated Handle and export APIs in the ros2_control framework. This change affects hardware_interface and controller_interface packages. Raw-pointer handles are replaced by shared pointers for data access. Legacy methods for exporting state interfaces and controller chains are removed. Users with builds warning from these interfaces may encounter build failures after the sync.
Remote pilots in command (RPIC) oversee fleets of delivery drones, acting as air traffic controllers for cargo delivery. This role requires aviation experience, FAA certification, and monitoring drone operations using GPS, lasers, and sensors. Unmanned aircraft systems traffic management (UTM) assists in collision avoidance for these autonomous drones. Pilots must maintain keen situational awareness to manage multiple drones and react to unforeseen hazards like sudden weather changes. Success in this field demands a collaborative mindset and the ability to remain calm under pressure.
A five-step workflow uses NVIDIA Omniverse libraries and Frontier AI models to convert CAD assets into simulation-ready OpenUSD for robotics. This process involves importing geometry, matching visual appearance, configuring physics properties, validating against SimReady requirements, and running pick-and-place tasks in Isaac Sim. Frontier AI assists by writing code to call Omniverse tools at each stage, reducing manual effort. The validation ensures that assets possess correct units, geometry, materials, and runtime behavior. The final step involves testing grasp and placement sequences with virtual objects, demonstrating successful robotic interaction.
Claire interviewed Kushant Patel from Brookhaven National Laboratory about robots used in scientific facilities like particle accelerators. Patel develops software enabling robots and instruments to perceive, map, and move in the physical world. He has experience teaching quadruped robots to inspect nuclear material and working on brain-computer interfaces. Currently, he is developing control systems for advanced synchrotron beamlines at Brookhaven. The Robot Talk podcast explores the field of robotics, AI, and autonomous machines.
Researchers at ETH Zurich transformed an off-the-shelf robotic hand into a self-contained robot that walks on its fingertips to manipulate the environment. This work demonstrates that body parts can perform functions beyond their original design, opening possibilities for new robotics. The team used reinforcement learning and novel penalty systems to train the asymmetric hand to crawl and navigate various surfaces. The mobile hand successfully demonstrated the ability to right itself, turn, and push objects autonomously in real-world tests. This research supports a vision for "autonomous modular embodiments" where robot parts can dynamically transition between different tasks.
Patent images show BYD's proposed humanoid robot design, featuring a bipedal structure with an oval face and articulated fingers. These renderings reveal a design with a narrow waist and exposed mechanical joints. BYD previously aimed to integrate humanoids into dealerships and eventually sell them for household use. Formal specifications and production plans for the robot remain unconfirmed at this time. A formal unveiling is planned for August at the Di Space exhibition venue.
ARC hosted Singapore's first full-sized humanoid fight night on October 7, featuring professional martial artists piloting Unitree G1 and EngineAI T800 humanoids. Teams Iron Fang and Steel Shadow competed in a survival-mode format, with pilots controlling the robots based on their body movements. Spectators could view the pilots' actions alongside the robot's health points in a split-screen display. The event highlighted the human-robot pairing as an early experiment in a new form of entertainment. This showcase contributes to a growing trend of robot entertainment formats that feature human interaction.
Dyna Robotics' Taku is scheduled to appear at the Nasdaq closing-bell ceremony on October 13. This appearance will follow a robotics founders forum in New York City. The forum will feature executives and demonstrations from companies like Dyna Robotics, Dexmate, and others. Taku, a semi-humanoid robot introduced with Dyna-2.1, is designed for connected workflows. The event is intended as an showcase rather than an announcement of new technical capabilities.
The Humanoid Hub Conference will be held October 15–16, 2026, at the AWS Builder Loft in San Francisco. Founder Devang Adhyaru emphasizes that in-person interactions with robot builders provide significant value beyond online content. The event aims to bring together researchers, entrepreneurs, and the robots they are building, focusing on research, deployment, and relatable robotics. Adhyaru expects widespread usefulness for home humanoids to be more than two years away, citing challenges in AI and achieving truly dexterous hands. The conference will feature talks, panels, and demonstrations from various robotics companies and research groups.
CRIS-0 is a robot AI system that combines a causal world model with an agent to predict and verify the outcomes of its actions. It functions by modeling the environment and predicting consequences, allowing it to replan or retry when reality diverges from its plan. The system checks execution results against predictions to catch errors before they propagate through a task. Aether emphasizes that this prediction-and-verification loop is crucial because physical actions cannot be reversed. The system aims to generalize beyond controlled demonstrations to handle new environments and tasks effectively.
Tesla's patent application details flexible touch sensors designed to conform to the curved surfaces of robotic fingers and palms. These sensors feature individually readable contact-sensing regions, utilizing conductive or resistive materials layered on flexible substrates. The filing extensively covers manufacturing methods such as printing and molding to achieve these three-dimensional sensor shapes. The application discusses creating a "humanoid tactile skin" and implies potential use in the Optimus robot. However, the document does not provide specific performance results or confirm the sensors' implementation in the Optimus hand.
A controlled study investigated whether dynamic-point filtering assists ORB-SLAM2 front-ends when texture is scarce in synthetic indoor scenes. The experiment varied surface texture and scene dynamics factorially across identical camera trajectories. Three filtering methods, including an optical-flow filter and a multi-view depth-consistency filter, were compared against the unfiltered ORB-SLAM2. Filtering primarily helped in the most dynamic areas, but for static scenes, filtering reduced tracking completeness and sometimes led to poor initialization. The results indicated that the benefit of filtering depended on the specific filter and the environmental conditions.
A method is proposed to teach legged robots diverse quadruped skills by combining Reinforcement Learning and Imitation Learning. The approach uses a three-stage process involving training multiple teacher policies on specific tasks. Subsequently, a student policy is trained using a multi-teacher distillation setup with a combined RL and IL objective guided by adversarial task selection. This method trains a student policy capable of performing 22 tasks, preserving motion quality and showing generalization to new tasks. The resulting policy was successfully deployed on a Unitree B1 quadruped.
TacHair is a framework that uses high-resolution tactile observations to guide online correction for robotic hair stroking and perception. It represents tactile data as a spatial hair-contact distribution to address the difficulty of regulating interactions with deformable hair. A visuotactile imitation policy generates the nominal motion, while a residual module separately corrects local contact deviations. Evaluations showed that this method improved both task progression and contact maintenance during hair stroking. The results confirm that spatial tactile contact distributions effectively aid contact-preserving interaction with visually occluded surfaces.
Masked Generative Motion Planning (MGMP) extends learned motion priors for structural repair, moving beyond local refinement. It utilizes a masked generative transformer to create discrete trajectory candidates in parallel. Geometry-Guided Token Search (GGTS) then uses scene geometry to efficiently search for the best edits and alternatives among these candidates. This approach enables route-level restructuring rather than just local trajectory deformation. The method achieved high success rates on various planning tasks and generalizes well to unseen scenarios.
TAPNAV is a tactile active-perception framework designed for humanoid navigation in vision-denied environments. It allows the robot to actively probe the surroundings to gather localization cues without relying on vision. The framework combines uncertainty-aware global route planning with information-gain-driven local probing. A whole-body controller coordinates locomotion and end-effector actions to execute the planned navigation and probing. Evaluations show TAPNAV achieves lower state estimation errors and higher task completion rates than existing methods.
NavGPT-3 is a harness that connects language models with physical action policies via an operating system-like runtime. This runtime manages reasoning, acting, and monitoring threads, allowing the robot to react to real-world events through interruption and thread switching. The system utilizes a NavGPT VLA action policy that allocates visual tokens based on scene changes. With the complete harness, NavGPT-3 achieves state-of-the-art performance in navigation benchmarks and approaches human-level performance in following paths. The integration significantly reduces reaction time from language-model decisions to action-policy steps, demonstrating a crucial connection between intelligence and physical control.
Variability exists in the outcomes of dynamic cloth manipulation even when the same trajectory is applied. This variability has not been systematically characterized, making it crucial for assessing the reliability of learned policies and simulations. Experiments involved running the same trajectory across various tasks with different cloth properties and speeds, yielding significant variability in marker positions. This variability is primarily driven by cloth physical properties and increases with motion speed. Furthermore, current cloth simulators fail to reproduce the magnitude and ordering of the variability observed in real cloth.
Long-horizon skill chaining often suffers from Observation-Space Shift (OSS) where performance degrades due to displaced scene states left by preceding skills. The dominant cause of this shift is displaced scene state, rather than robot configuration or manipulated objects. A proposed solution is a detect-restore-resume system that monitors for stalls, restores scene components, and allows the skill to resume. This system significantly improves full-chain success on the BOSS-44 benchmark, outperforming other methods. Results suggest scene restoration is a more effective strategy for long-horizon composition failures than simply retrying from an off-support state.
Skill-SLM proposes a framework for reliable robot operation by treating it as a task-decomposition and skill-composition problem for Small Language Models (SLMs). It uses a robot operational skill-aware context-free grammar to create a skill library from natural language task instructions. LLM teachers are used to synthesize training data, enabling SLMs to reliably decompose tasks and orchestrate skills. A progressive skill orchestration strategy further enhances the reliability of skill implementation. Experiments on UAV and ground vehicle tasks show Skill-SLM significantly outperforms distillation-oriented baselines, especially on unseen tasks.
The Latent Action-Conditioned Robot World Model (LAC-WM) introduces a unified latent action space to address the challenge of building robot world models that generalize across different embodiments. This unified space improves world model performance when adapting to novel robot embodiments compared to an Explicit Action-Conditioned World Model (EAC-WM). Evaluation on dexterous manipulation and LIBERO benchmarks shows LAC-WM outperforms EAC-WM by significant margins. The unified action space allows LAC-WM's performance to scale positively with the number of pretraining embodiments. These results emphasize the benefit of a unified action space for efficient cross-embodiment learning in robotics.