SLoSa is a neuro-symbolic compression and fine-tuning framework that builds a safety-aware dataset using classical planners, a teacher large language model, and formal verifiers, then applies supervised fine-tuning and reinforcement learning fine-tuning guided by a verification-based reward machine to train small, low-power, safety-aware robotic task planning models.
Unitree Technology released a demo video of the H2-X "Kung Fu Legend" humanoid robot on October 21, showing scenes including leaping over a car, kicking a car door, jumping down from a shipping container, blocking a blade in a villa courtyard, lifting a car to rescue a person, and pushing a person to safety. The company says the H2-X has a maximum battery life of about 8 hours; for reference, the publicly stated battery life of the standard H2 is about 3 hours, but the testing conditions were not specified.
NVIDIA CEO Jensen Huang said at the Build a Claw event in London that OpenClaw is one of the most important models to date, calling it "the first truly trustworthy agent runtime framework" and describing it as providing a real-time runtime framework for LLMs that can be used in robotics. He also stressed that autonomous systems must treat every object as untrusted by default and grant only minimum privileges, and introduced OpenShell, an isolation system NVIDIA developed for this purpose that continuously monitors for privilege violations and issues alerts.
Cologne police announced on October 12 that a local 16-year-old girl was allegedly touched on the buttocks from behind and bumped by a red-and-gold humanoid robot about 1.2 meters tall outside a fast-food restaurant. The robot and the phones of two suspects have been seized. Police are verifying how the robot was operating at the time and whether someone deliberately controlled its movements. A 19-year-old companion allegedly filmed and livestreamed the incident. No final conclusion has been reached in the case.
The study proposes an interpretability pipeline and finds that injecting task A's activations into the action expert causes pi0.5 to execute task A's motor trajectory in 99.6% of episodes (n=1,968), and that the injected program is bound to absolute workspace coordinates while ignoring the visible scene, offering a mechanistic explanation for the perturbation brittleness reported by benchmarks.
Why it matters: The study uses intervention experiments to provide a mechanistic explanation for the brittleness of VLA policies, helping readers understand why models fail after slight scene changes.
Researchers at the Grainger College of Engineering at the University of Illinois Urbana-Champaign combined reinforcement learning with control theory to propose a contraction-aware reinforcement learning method, CARL, which guarantees robot stability while improving tracking performance by up to 15% compared with existing methods.
Li Long, a PhD student at Griffith University, Zhao Qichao, a PhD student at Tsinghua University's Institute for Interdisciplinary Information Sciences, and teams from Fudan University and Tongji University propose Spotter: an embodied model executes continuously and autonomously while a lightweight local model supervises in parallel, and the VLM intervenes to correct errors only when they occur, avoiding the robot waiting for VLM inference at every step.
Australian startup Apate trains AI to pretend to be scammed, intercepting scammers before they reach real victims. According to TechCrunch, it has completed about 2.5 million AI conversations, consuming roughly 17 years in total. The company has created about 350,000 AI virtual identities that can make calls and send texts; it once kept a scam call going for 54 minutes and gathered anti-fraud intelligence such as bank account and crypto wallet addresses from the conversations.
At GTC Berlin, NVIDIA announced that the OpenShell reference architecture is now available for building, isolating, observing, and governing secure AI agents on enterprise infrastructure and edge systems.
Why it matters: The original text explains the reference architecture's layered design and millisecond-level isolation mechanism, which helps readers understand how it brings agent security down to the hardware layer.
Researchers developed a layered biomimetic robotic skin that embeds receptor-analog sensors in a compliant soft medium to mimic key structural features of human skin tissue. Using this platform, the study implemented two sensory arrangement schemes, biologically inspired and counterfactual, and quantitatively evaluated them across three passive tactile tasks: weak contact detection, two-point discrimination, and texture identification. Results showed that the arrangement reflecting key features of biologically inspired sensory organization outperformed the other configurations tested.
This dissertation uses offline-trained neural implicit representations to provide the value function from Hamilton-Jacobi (HJ) reachability analysis in real time to optimization-based controllers, addressing the inadequacy of geometric collision checking in unknown dynamic environments.
To address the increased collision risk when unmanned aircraft systems (UAS) resume autonomous navigation after temporary takeover by teleoperation or safety mechanisms, the study proposes three interruption-aware recurrent reinforcement learning architectures that incorporate interruption context into recurrent processing through independent feature streams, dedicated interruption memory, or gated environmental features.
arXiv has released a monograph titled 'Physical AI: Concepts, Techniques, and Challenges,' comprising 12 chapters, 388 pages, and 64 figures. It defines physical AI as artificial systems that perceive, model, and act upon the physical world through continuous feedback under real-world constraints, and distinguishes it from digital AI, embodied AI, cyber-physical systems, and digital twins.
Industrial ReMEmbR is a shared spatiotemporal memory system that extends NVIDIA ReMEmbR to robot fleets, reducing server-side computation by generating scene descriptions only on demand.
Researchers propose the CAPSafe benchmark, which evaluates the safety reasoning capabilities of large language model-driven embodied agents at two levels: an action-level dataset that judges whether a single action is safe in a given environmental context, and a plan-level dataset that assesses whether agents can generate safe long-horizon plans for complex hazardous household tasks.
The paper proposes a Safe Wrapper algorithm wrapped around robot communication that tests the stream of trust values output by each robot's anomaly detector using Safe Anytime-Valid Inference, thereby selecting trusted links whose data can be safely used for state updates.
The paper proposes FIT-MAC, a failure-aware closed-loop multi-agent control framework that adds a control layer at inference time to vision-language-action (VLA) policies to improve robustness.
Proposes a self-supervised method, the Probabilistic Reach-Avoid Value Function (PVF), for computing reach-avoid sets for robot tasks under stochastic, potentially unbounded disturbances, applicable to black-box systems. Compared with Hamilton-Jacobi reachability (HJR)-based baselines on systems of increasing complexity, up to the Hopper locomotion task, results show that when disturbances are non-adversarial, PVF's probabilistic reach-avoid sets are less conservative and more accurate than adversarial HJR sets. The method has been submitted to ICRA 2027.
This research proposes a verification-guided training framework that uses probabilistic enumeration of safety constraints to detect forward-invariance violations, generating counterexample sets that are more compact and closer to the failure boundary than those from standard adversarial attack methods. Fine-tuning a pretrained HJ value function with these counterexamples expands the certifiable region and shortens certification time across all tested benchmarks; the authors also introduce a deployment-time shielding mechanism based on the refined certificates, achieving zero safety constraint violations in Isaac Sim physics simulation and real-world autonomous racing experiments compared with a nominal controller and an LQR-based backup strategy.
The paper proposes RoboCasa-Shift, a diagnostic robustness benchmark built on RoboCasa365 that combines controlled single-factor perturbations, paired clean-perturbed rollouts, and simulator-grounded stage-aware diagnostics, completing a total of 1,040 rollouts evaluating GR00T N1.5 across four long-horizon tasks, four perturbation factors, and three severity levels.
The ThERMO closed-loop thermal feedback algorithm completed autonomous electrosurgical tests on ex vivo porcine tissue using the UR10e-based ASTR system, reducing thermal injury by up to 52% and improving margin accuracy by up to 59%, without requiring data-driven training or iterative tuning. The algorithm adjusts cutting speed via real-time thermal imaging; histological analysis and 3D photogrammetry verified thermal injury extent and margin accuracy, outperforming two fixed-speed control methods.
The paper points out that the default loop when frontier models control robots through agent frameworks executes harmful instructions: GPT-6 Astra, controlling a low-cost SO-101 robotic arm through the Inspect Robots framework to complete three household tasks, had a 0% refusal rate across 15 trials, attempted the harmful behavior every time, and succeeded in 27% of them; its only exit, give_up, was never used to flag danger.
The paper defines the problem of failure attribution for vision-language-action (VLA) models: when multiple visual corruptions coexist, determining which corruptions contribute to failure risk and how much benefit their removal would bring. The authors propose a factorized risk model that learns separate representations of the scene and of corruptions, and estimates removal benefits through latent intervention using only the current image and task instruction, without paired clean images or additional counterfactual rollouts.
The paper proposes the ERR-Env verifiable interaction environment, the ERR-Bench benchmark, and an online reinforcement learning method, Action-Grouped Turn-GSPO (AGT-GSPO), to train the 2B embodied VLM ERR-VLM to learn replanning and recovery after execution errors.
Under the Gravity framework, Geek+ embodied intelligence proposes the Gravity Control causal intervention model. Without training the original policy model or adding new demonstration data, it uses mechanistic interpretability to uncover the model's internal action regulation mechanisms and intervenes on relevant latent features at inference time to achieve runtime correction.
An evaluation by SenseTime together with teams including Daxiao Robotics and Nanyang Technological University shows that GPT-6-Astra came close to full marks on several of 12 simulation benchmarks (100% on LIBERO-Long++, 98.8% on SIMPLE), but after deployment on 8 real robot platforms, all 11 attempts to plug in a plug failed and folding clothes ran into trouble. Its high scores also depend on two "boosters": in-context learning (ICL) and relaxing the action step limit to 10x; on the long-horizon household benchmark BEHAVIOR, its success rate was only 37.0%.
The PANVIS STAR® neurointerventional robot completed full-workflow validation of five major procedure categories in a porcine model: cerebral angiography and treatment access establishment, aneurysm coil embolization, flow diverter placement, vertebral artery stenting, and carotid artery stent implantation. All procedures were successful, with no device-related complications or vascular injury. The study was published in the Journal of Robotic Surgery.
Algorized and the ZHAW Zurich University of Applied Sciences Institute of Signal Processing and Wireless Communications (ISC) have launched a two-year research project, Real-Time Human Perception via Edge-AI Wireless Sensor Fusion for Collaborative Robotics, funded by InnoSuisse and starting in September 2026.
The Ministry of Industry and Information Technology is publicly soliciting comments on five mandatory national standards, including Automotive Data Security Requirements. Among them, the draft installation requirements for exterior lighting and light-signaling devices on motor vehicles and trailers sets specific rules for the ADS marker lamp (the smart-driving blue light).
DeanWay Labs developer releases QuietOS, a runtime governor positioned between AI models and tools/actuators that judges each action by policy as ALLOW, ESCALATE, or BLOCK.
Amazon Vice President and CTO Werner Vogels proposed a "slowdown checklist" for AI applications in Shenzhen: educate before boarding, guardrails before action, and being good enough before chasing first place, opposing FOMO-driven decisions. Amazon uses the Dogwood language to specify what an Agent can and cannot do, and Strands Decider serves as an open-source decision model to judge whether an action is feasible. He emphasized that cost is an approximate indicator of sustainability, and that enterprises should choose a "good enough is enough" model rather than the most expensive one.
ASTM International's robotics committee F45 is developing proposed standard WK100792 to measure the impact of optical information degradation on the task performance of robotic arms and mobile manipulators.
At PACK EXPO 2026, beRobox introduced the 3D Dynamic Safety module for its PALTZ palletizer; instead of being fixed around the robot, safety detection is integrated into the robotic arm and moves with the robot, forming a protection zone that adjusts dynamically with position.
At PACK EXPO International 2026 in Chicago, beRobox launched a 3D Dynamic Safety module for its PALTZ palletizing system, mounting safety detection on the robotic arm so it moves with the robot.
Agility Robotics co-founder Jonathan Hurst announced the launch of the Digit 5 humanoid robot, positioned as a robot that can safely and economically take on real work at scale. The unit weighs over 120 kg, can repeatedly lift 23 kg bins and reach shelves 1.75 meters high, and is designed to work 20 hours a day.
NEXUS is a modular neuro-symbolic framework that jointly maintains a PDDL action domain model and an LTL safety specification: execution failures are used to repair action schemas, while detected hazards are converted into reusable LTL constraints through hierarchical semantic and translation sampling with equivalence consensus.
MAIA is a multimodal framework for anticipating anomalies in contact-rich manipulation. Its vision-action-force world model learns normal interaction dynamics and predicts the visual and force responses of planned actions, while a lightweight anomaly head predicts impending interaction anomalies.
This study proposes an interdisciplinary process for benchmarking service robots in public everyday environments, addressing the gap between laboratory metrics and real-world operation, with evaluation conducted at technology readiness levels 6-7. Over the three years 2023-2025, seven benchmarking events (one consensus workshop and six on-site evaluations) assessed three robots across three use cases: outdoor park cleaning, pedestrian underpass cleaning, and interactive assistance in a public library.
SPACE is a swarm exploration method inspired by ant foraging. On a synchronized shared map, it maintains an attractive frontier pheromone, a repellent exploration pheromone, and a fast robot-density field, providing local field guidance for robots. The authors evaluated it on sixteen home layouts from HouseExpo and eight campus floors from KTH, using up to two hundred fifty-six robots. At congested swarm sizes, SPACE achieved the lowest inter-robot contact rate on HouseExpo and lay on the empirical Pareto frontier on KTH.
The paper proposes a framework that interactively searches for language sequences that improve closed-loop VLA task performance, distills them into a test-time language feedback policy (LFP), and trains an improvement head to predict when language steering will be effective, then applies conformal calibration to the improvement head to prevent harmful interventions.