World models aim to capture how the physical world changes when an agent takes an action. These models help systems predict future states based on current observations and dynamics. Approaches vary, including visual-token models, latent dynamics models like Dreamer and JEPA, and object-centric models. The ultimate goal is to enable agents to reason about and interact with the physical environment effectively. Core patterns involve learning the dynamics that govern transitions between states.
Information is being shared following the meetup on Kubernetes/KubeEdge with ROS 2 at ROSCon Toronto. The author invites readers to subscribe to the thread for further details. A follow-up online meeting is planned after ROSCon, possibly in October. The author will send an online meeting invitation once they return to Tokyo. This communication serves as a follow-up to the discussions held at the event.
rclrs v0.8.0 has been released, supporting ROS 2 Lyrical Luth bindings for multiple ROS 2 distributions. ROS message packages are now accessible via ros-env to unify message definitions across rclrs and applications. ActionClientState now provides server availability, and parameter callbacks are supported. Performance improvements and fixes were included across action handling and ROS initialization. Applications using generated messages must now import them with `use ros_env::*;`.
ROSCon Global 2026 is currently live from Toronto. Streams are accessible on the ROSCon website. The primary live stream is hosted on Vimeo in the Grand Ballroom West/Centre. A secondary stream is also available on Vimeo in the Grand Ballroom East, starting after lunch. Each post corresponds to one participant.
Offloading physical AI inference to edge or cloud GPUs offers significant advantages over relying solely on onboard robot GPUs. This offloading improves task success rates, enables larger AI models, and enhances robot response in dynamic environments. Offloading also substantially extends robot battery life by reducing the power consumption of onboard AI compute. A new toolset, leveraging Kubernetes, facilitates the containerization and orchestration of distributed robotics AI workloads. This capability is integrated into the Physical AI Toolchain, enabling scalable and real-world deployment of physical AI.
IROS 2026 will be held in Pittsburgh, USA, from September 27 to October 1, featuring plenary talks, workshops, and forums. Keynote speakers will address topics ranging from autonomous surgery to embodied AI and soft robotics. Community discussions will cover subjects like US government funding and robotics education. Plenary panels will explore themes such as start-ups, space, and the role of robots as generalists or specialists. The event will also include numerous workshops and tutorials.
PickNik Inc. and Qualcomm Technologies, Inc. are merging, with the move driven by the significance of the MoveIt project in robotics. MoveIt has a vast community, contributing code, documentation, and enabling real-world robotic applications through motion planning. The future of robotics involves a hybrid approach combining AI models with traditional MoveIt methods for reliable deployment. Qualcomm is acquiring PickNik to address the need for integrating advanced AI models directly onto physical hardware for real-time robot operation. This collaboration is expected to continue supporting the open-source development of MoveIt and ROS.
Navigating off-road terrain is difficult because traversability is complex and hard to label at scale. FLINT is a lightweight traversability estimator with a 21.6M-parameter backbone, significantly smaller than foundation models. It achieves accurate costmaps with fewer sensors and less computation than deployed foundation-model systems. Experiments show that heavy sensing and computing are not necessary for traversability estimation. The best self-supervised models achieve near-full autonomy in field trials.
Learning-based models used for industrial robotic arm manipulation present security risks due to the direct link between model predictions and physical actions. Backdoor attacks can cause the robot to perform unintended actions only when specific triggers are present, even during normal operations. The study empirically investigated backdoor attacks and defenses on commercial robotic arms (FANUC and xArm) to test for stealthiness. An online defense pipeline was developed to detect and neutralize triggers at runtime, which was compared against an offline fine-tuning defense. The research also assessed the computational overhead introduced by the defense mechanism for industrial applications.
MSK-Bench benchmarks full-body musculoskeletal motor control across 22 tasks categorized into postural stabilization, locomotion, and environmental interaction. It evaluates five control paradigms, including RL and imitation-based methods, using unified protocols and robustness perturbations. The benchmark reports task success alongside physiological diagnostics like activation cost and joint smoothness. Findings indicate that embodiment-aware exploration and imitation priors offer benefits, but improved task success does not always correlate with better physiological agreement. MSK-Bench establishes a testbed for evaluating full-body muscle-actuated control based on tasks, methods, and metrics.
Event cameras offer low-latency perception, but existing detectors suffer from delays that make predictions stale for real-time robotic decisions. The study investigates this observation-availability mismatch in event-based multi-object detection. ChronoFuse is introduced as a causal availability-time detector that predicts object states for output availability rather than observation time. This method recovers significant accuracy lost due to latency, nearly restoring zero-delay performance under extreme motion. The research demonstrates that predicting ahead is crucial for fast-changing environments like autonomous driving and agile flight.
Spiderbot is an open-source, energy-efficient hexapod robot designed with passive gravity compensation. It employs a 4-bar linkage with a passive spring in each leg to mechanically support body weight, significantly reducing torque requirements. This mechanism consumes only 1.5W during the standing stance, resulting in energy savings of over 90% compared to unsprung versions. The platform supports payloads up to 3.25kg and is suitable for large-scale experiments within a modest power budget. The system's locomotion capabilities were validated using a Reinforcement Learning policy and demonstrated on various terrains.
A fly-inspired recurrent controller is investigated for humanoid locomotion. This controller couples 3,609 continuous neural states to a simulated robot using body-observation projections and motor-neuron-labelled readouts. The system is evaluated across various terrain instances, speeds, and yaw offsets in a simulation study. The research demonstrates that the embodied recurrent control system supports locomotion through direct body-and-command input and carried motor state. The findings provide a foundation for comparing circuit structures and control resources in locomotion.
The research presents an end-to-end framework to learn expressive humanoid locomotion from monocular runway videos for robot fashion shows. This framework involves motion recovery, robot retargeting, motion correction, policy training, and simulation-based evaluation. The method transforms runway videos into deployable locomotion policies suitable for humanoid robots. Evaluation on the Booster K1 robot showed the learned policy successfully performed stable and expressive catwalk motions without falling. The results indicate the framework's potential for applying expressive motion capabilities to humanoid robotics in performance settings.
MultiPush is a reinforcement-learning framework designed to learn efficient pushing schedules and robot allocations for rearranging objects in a constrained workspace using car-like pushers. The method exploits the kinematics of the domain by rephrasing the problem as an ordered assignment of Dubins curves to the robots. It achieves performance improvements by using a constraint-aware traversability graph to guide the learning process. Simulations show MultiPush reduces makespan by up to 16% compared to baselines while requiring faster planning time. The method was demonstrated successfully in a real-world scenario involving the rearrangement of twelve objects by two or three robots.
A laser-tracker-assisted method is presented for calibrating camera-equipped mobile robots. This technique integrates laser-tracker-based 3D metrology with camera-based 2D observations. The method provides a generalized formulation for calibrating the camera pose relative to tracker-localized mobile robots. It achieves a more general approach by chaining multiple calibration targets, relaxing previous assumptions about robot and camera configuration. This allows the method to support various camera-equipped mobile robot systems.
HiRE is a training-free framework designed to overcome the reward bottleneck in policy finetuning using hindsight reward editing. It aligns foundation representation models with physical control awareness by contrasting successful and failed trajectories. The method identifies and adjusts for "trap states" that are mispredicted as rewarding or failed, explicitly penalizing traps and boosting critical successful states. Experiments demonstrate that HiRE provides dense, control-aware feedback, leading to superior sample efficiency and performance compared to other reward recipes. This results in stable policy updates and higher performance ceilings for reinforcement learning.