Skild AI bets that robotic intelligence should be transferable across different hardware rather than being rebuilt for every machine. Their research focuses on connecting motor adaptation, video-prompted task learning, and self-play to achieve this goal. The core idea is to separate intelligence from the physical machine so that understanding a body is learned by the model. The company aims for the model to learn body understanding rather than requiring engineers to specify every detail perfectly beforehand. The ultimate test is whether these capabilities can survive the real-world requirements of speed, reliability, and safety in deployment.
Google's Universal Cart is now available in the U.S. across Search and the Gemini app to simplify holiday shopping. It allows users to save products from multiple retailers into a single place, eliminating the need for multiple browser tabs. The cart actively tracks price drops, identifies relevant promo codes, and suggests payment methods. Users can complete purchases securely using Google Pay or transfer items directly to the merchant's site. This tool aims to give shoppers confidence and control over their online shopping experience.
Industrial water reuse involves treating effluent to be returned as process water, which can reduce fresh water consumption and discharge volume. Successful reuse requires a four-step process: diagnose water, engineer treatment, pilot the system, and deploy the full-scale system. Automation systems instrument the water loop using probes for conductivity, pH, and ORP, with a PLC controlling chemical dosing and water diversion. Recovered water is used for various applications like cooling tower makeup, washdown, or process feed, each with specific quality requirements. Implementing this system rewards the same control discipline used in production lines by measuring real input and proving performance before scaling.
Claire discussed using artificial intelligence to develop complementary robot hardware and algorithms with Rika Antonova from the University of Cambridge. Antonova's research interests include co-design, automated algorithm discovery, and reinforcement learning in robotics. She has a strong academic background from institutions like Carnegie Mellon University and Stanford. Robot Talk is a weekly podcast exploring the fields of robotics, AI, and autonomous machines. The podcast features discussions on various aspects of robotics research and application.
The next Cloud Robotics WG meeting is scheduled for Monday, October 19, 2026, at 4:00 PM UTC. The primary focus of this meeting will be continuing the editing of the first draft of the Logging and Observability guide. The group aims to complete the guide's edits during this session and submit it to guest speakers for review. Information regarding the meeting link and notifications can be found via the provided links. Attendees are encouraged to join to participate in the guide revision process.
Tesla filed a US patent for a touch-sensitive "skin" designed to mold around robot fingers and palms for detecting contact, force, and pressure. The patent describes a manufacturing process where tactile sensors are printed onto flexible materials and then heat-formed into three-dimensional shapes. The proposed sensor utilizes overlapping conductive regions on flexible substrate layers to register pressure changes through capacitive and piezoresistive methods. The technology emphasizes manufacturability using established techniques like screen printing and thermoforming for scalable production. This innovation aims to advance humanoid robots beyond basic grasping by enabling them to feel how firmly they are manipulating objects.
Skild AI developed a self-play training method using its S1 model to teach robots physical soccer skills through simulated competition. The robots learned complex behaviors like dribbling and tackling by playing against increasingly capable versions of themselves, rather than receiving explicit behavioral instructions. This method allowed the robots to develop skills from a single objective, leading to capabilities that emerged from the competitive process. After 140 years of simulation, the learned policy was successfully transferred to a physical humanoid robot. This approach aims to reduce the reliance on human demonstrations for teaching robots complex physical skills.
ETH Zurich scientists demonstrated how a WUJI robotic hand can achieve self-propelled locomotion and manipulation without a separate mobility system. The research trained the hand to use its fingers for crawling, recovering from falls, and moving around its environment. This "compact mobile manipulator" was capable of performing self-supported tasks like pressing keyboard keys and pushing objects. Reinforcement learning was used to develop the locomotion, taking into account the unique characteristics of each finger. This work illustrates an unconventional approach by integrating locomotion and physical interaction capabilities into a single robotic hand.
Pudu Robotics is expanding its focus beyond service robots into developing semi-humanoid and quadruped robots. At IFA 2026, the company showcased these advancements, including the D7 semi-humanoid and the D5-W quadruped robot. The D7 focuses manipulation in its arms, waist, and head, capable of lifting up to 20 kg. The D5-W robot combines wheels and legs for varied terrain navigation and can autonomously recover from falls. This diversification extends Pudu's offerings and positions the company in competitive areas of industrial robotics.
Faraday Future and FFR have launched the $2,490 FF Nova humanoid for schools, initially releasing only 100 invitation-only units. The Nova is a research and teaching platform measuring 49.7 cm and capable of walking, turning, and recovering from being knocked over. It features various sensors and supports ROS 2, with an additional software package available. A smaller, $299 Nova Mini is available for children, supporting visual block programming. The launch is part of a broader restructuring of the robotics business involving an acquisition.
PNDbotics' full-size Adam humanoid can climb onto a one-meter high box. This capability was detailed in a LinkedIn post by Space and Technology. The process involves the robot analyzing the obstacle, planning movements, and placing one foot on the box before shifting its weight. The climb requires the robot to support and raise its body while maintaining balance. The movement involves coordinated action across the entire body, with weight transfer being central to the ascent.
Waymo is launching its fully autonomous ride-hailing service in Toronto, aiming for first riders in 2028 through a phased, safety-first approach. The rollout will involve mapping local street networks and testing with autonomous specialists to ensure adaptation to Toronto's driving conditions. The expansion involves a partnership with Element Fleet Management to create jobs in fleet management, validation driving, and customer service. The Waymo Driver has been associated with significantly fewer serious accidents compared to human drivers based on autonomous driving data. The service aims to complement Toronto's existing transportation ecosystem while prioritizing safety and accessibility.
Causal analysis of vision-language-action (VLA) models investigates how perceptual and semantic representations are transformed into actions. The study found that causally relevant information influences the action expert through a specific "action-conditioning interface" in mid-to-late cross-attention layers. Restricting the action expert conditioning in one model reduced inference FLOPs by up to $33\%$ while maintaining task success. This principle was also demonstrated for real-robot deployment of other models. The research suggests that causal understanding can be directly used to improve the efficiency of robot foundation models.
Multi-dimensional featurizers, known as block-sparse featurizers (BSFs), are investigated to recover interpretable structure in robot foundation model features. These featurizers successfully capture task-relevant features such as perception, language, and proprioceptive state underlying the model's computation. Fine-grained steering experiments demonstrate that manipulating these features causally affects downstream action predictions. Furthermore, feature-based data attribution allows for isolating the specific training tokens responsible for observed behaviors like failure recovery. Overall, matching featurizers to representation geometry provides a more effective method for understanding and controlling robot behavior.
The paper establishes feasibility ceilings for drone-mesh interception before any learning policy is chosen. It provides closed-form bounds for performance based on the spread and timing of agents covering a predicted arrival point. Analysis shows that naive area ratios miss coverage ceilings, indicating constraints from separation, not the problem itself. Optimal packing is achieved by hexagonal arrangement, and other configurations yield less coverage. Analytic placement achieves 85% of the configuration's potential, while learned policies achieve significantly less, suggesting these bounds should guide learning experiments.
Persistent profile-guided reset-free recovery (PGRR) is introduced to improve final gait selection after irreversible leg damage. This method predicts multiple outcomes over repeated gait execution and uses uncertainty-aware estimates to guide short online probes without resetting the robot. Across 108 matched tasks involving leg amputation, fracture, and joint lock, PGRR improved final quality by 0.106 and slot-averaged recovery quality by 0.044 compared to a baseline. Furthermore, it reduced low-quality failures by 27.8 percentage points in all damage classes. Conservative uncertainty handling was identified as the clearest contributor to these performance gains.
A methodology is presented for adaptive mission planning for solar-powered autonomous surface vessels. This strategy maximizes the information-driven quantity of clarity, which is mathematically linked to differential entropy. The approach uses a clarity-driven ergodic search to adapt a target clarity map based on in-situ measurements. An ergodic path planner continually re-plans considering the target clarity map and mobility restrictions from the solar power source. Simulations at Jordan Lake demonstrate the effectiveness of this approach for such ASVs.
Existing safety systems fail when autonomous vehicles lose control due to events like icing or hydroplaning, as they focus on preventing instability rather than recovery. The research trains a racing policy using reinforcement learning to learn drifting as a time-optimal maneuver, then adapts it for recovery with a minimum-time reward. The method defines recovery based on a stable equilibrium and uses Gym-Khana for controller training and evaluation. This transferred policy significantly improves recovery over baseline methods and outperforms other controllers. The policy, when deployed on a physical vehicle, demonstrated the ability to recover from states the baseline could never handle.
Sustainably powered mobile robots can maximize information quality by optimizing their speed trajectories while managing energy. This involves solving a persistent, infinite-horizon speed trajectory optimization problem posed as a periodic optimal control problem. The study uses information-theoretic clarity as a metric to derive necessary conditions for optimal control, leading to follow-on properties characterizing the optimal speed choices. The research also establishes a high-speed information limit as energy reserves approach infinity. Simulation results, based on salinity characterization in the Chesapeake Bay, demonstrate the practical relevance of these optimized trajectories.
The research introduces BubbleCBF, a constructive method for designing feasible safety constraints for second-order control-affine systems. This involves constructing collision-free spheres, or "bubbles," from a signed distance function to define a velocity-dependent safety constraint. A quadratic-program safety filter is used to enforce the resulting safe set, which is represented by the overlap of these bubbles. The method proves the resulting controller is well-defined, continuous, and forward invariant under mild assumptions. The approach is demonstrated to achieve safe autonomous navigation for a differential-drive robot in unknown environments.