Mobile manipulators and humanoids have emerged as a growing field in robotics, with some already being deployed in factories and warehouses. Industry experts discussed technical concerns such as power management, perception, and data collection for AI and world models, with panelists sharing lessons learned from scaling advanced systems from the automotive industry. Lattice Semiconductor emphasized the need for reliability and security in AI-enabled humanoid robots, particularly those used in public-facing applications. Researchers are also exploring human-machine interaction and skills training for humanoid service robots.
Researchers and industry professionals at the Imperial Robotics Summer School, hosted at Imperial College, London, explored challenges of multi-party human-robot interaction. The group focused on enabling robots to hold conversations with multiple people simultaneously, detecting active speakers in real time, tracking individuals through interactions, and generating natural gaze behavior. This project highlighted the complexity of multi-party human-robot interaction and the importance of interdisciplinary research in robotics.
Companies like 1X and Tesla are investing in tendon-driven hands for compliance, while Sharpa and Wuji Tech favor direct-drive motors with high precision. Figure's CEO Brett Adcock recently criticized the company's early tendon-driven hand design, calling it an "engineering mistake," but Foundation engineer Andrea Esposito disagreed, citing a "skills issue" related to poor routing and weak actuation. The debate highlights the ongoing struggle to develop a truly general-purpose humanoid robot with a reliable and effective hand.
ROS Discourse General: Launching VLAs on the Humanoid robot (G1, T800) for continuous tasks is currently hindered by the realsense pack of cameras, particularly requiring multiple cameras for implementations like the unifolm vla, which necessitates 2 cameras on wrists and 2 on shoulders.
Northrop Grumman has launched the Mission Robotic Vehicle (MRV), a robotic spacecraft designed to service satellites in orbit, capable of satellite inspection, relocation, repair, upgrades, debris removal, and robotic assembly. The MRV features two fully articulated robotic arms that can manipulate specialized tools while conducting servicing operations. It will be able to install propulsion modules on existing satellites, extending their life without replacing the entire satellite. The spacecraft itself has a passive refueling module and computing hardware intended to support future AI-powered operations.
AutoStore has signed a global supply agreement with Amazon, establishing terms for future procurement of its automation systems. The agreement does not commit Amazon to purchasing any particular quantity of equipment and no financial terms or potential order volumes were disclosed. AutoStore says the deal creates an established framework for future installations without requiring negotiation on underlying terms for each procurement. Approximately 2,000 of its systems are currently installed across 68 countries, with a focus on "Intelligent Fulfillment" combining automation, software, and AI to manage and optimize fulfillment operations.
Schaeffler is expanding its humanoid robotics product portfolio with formed strain wave gearboxes, key components in actuators that make up around half of the total costs of manufacturing a humanoid. The company's innovative forming technology can manufacture these components in seconds, cutting production time by minutes and reducing material consumption by over 75%. This method achieves comparable torques and efficiencies to conventionally manufactured components while offering high dimensional accuracy and process stability. Production will start in Germany in 2027 with plans to roll out to other regions.
Geekplus launches RoboShuttle Hyper, its next-generation tote-to-person climbing AMR solution for high-throughput warehouse fulfillment. The system delivers up to 6,000 totes per hour per 1,000 square meters and features parallel horizontal and vertical climbing to enable 50% higher concurrency than the industry norm. It also supports dual-robot operation within the same column and has a column-rail-integrated system for improved reliability.
Sony has filed a US patent application for a robotic hand featuring a human-like opposable thumb, providing evidence of the electronics giant's renewed research into humanoid and embodied robotics. The application, published by the US Patent and Trademark Office on August 6, describes a robot hand designed to reproduce the movements that make the human thumb important for manipulating objects. The patent focuses on the mechanical aspect of the hand mechanism, which breaks movement into three separate rotation points to enable opposition between the thumb and fingers. This design could have applications beyond humanoid robots, including service robots, remotely operated machines, and systems using tools designed for humans.
LG is developing a next-generation bipedal humanoid robot powered by Nvidia Isaac GR00T, with a public unveiling planned for the first quarter of 2027. The robot will be built using Nvidia Jetson Thor for onboard compute and advanced reasoning. LG plans to deploy wheel-based robots, including the CLOiD model, in its Tennessee washing machine manufacturing line and other global production facilities, homes, and commercial spaces.
Robot.com has unveiled R-dog, a four-legged delivery robot designed for navigating the last 100 meters of deliveries, particularly in areas with obstacles such as stairs, curbs, and uneven ground. The robot can carry large payloads and features an onboard compartment that can hold multiple items at once. R-dog's second application is interactive advertising through R-ads, which allows brands to remotely upload and manage campaigns on the robot's dual video screens.
RoboSynChallenge: Mastering Real-World Dexterity via Generalizing Synthesized Manipulation Skills introduces a unified benchmark for evaluating and advancing robotic manipulation policies across various tasks, environments, and difficulty levels. The challenge integrates large-scale synthetic data generation with standardized real-world robotic evaluation to foster the development of broadly capable, data-efficient, and adaptable manipulation systems. Participants are encouraged to leverage synthesized state-action trials to improve general-purpose policy learning, while final assessments are conducted exclusively on unseen real-world manipulation environments.
Researchers have developed Attune, a self-annotation tool to understand robot operator attention profiles. The tool automatically identifies meaningful gaze shifts and provides AI assistance for annotating why shifts occurred. It outputs a summary of operator gaze patterns for review. A user study revealed variation in observed gaze patterns and how Attune characterizes operator attention.
FUSE, an adaptive semantic-geometric evidence acquisition framework, combines explicit uncertainty-driven exploration with a learned amortized planner to efficiently select informative viewpoints for embodied agents. This approach enables active functional affordance grounding, where an agent sequentially explores a scene to identify and spatially ground an object satisfying a functional query. FUSE achieves the highest observed non-oracle grounding performance while reducing computation by 1.33x relative to fully explicit exploration.
SAP-Nav, a fully online and zero-shot framework, addresses the challenge of hierarchical open-vocabulary object navigation (OVON) through active perception. It constructs a Queryable Spatial-Semantic Representation from actively acquired room views, enabling spatial semantic queries from any explored location. SAP-Nav employs Active Viewpoint Verification to assess observation evidence and reposition the agent for more informative viewpoints before verifying candidates against category and attribute constraints.
This paper proposes a decentralized approach using a genetic fuzzy system to coordinate multiple robots in planetary missions. The system minimizes the total path length while avoiding obstacles by analyzing an elevation map and converting it into a traversability map. A genetic algorithm optimizes fuzzy inference systems for transporting objects, which are then validated in different testing environments.
AirForesight, a current-to-future spatial map imagination framework for UAV-VLN, learns a structured current-map representation from multi-view observations and jointly supervises it with future-trajectory prediction. This framework uses structured causal attention to propagate current spatial knowledge to future-map reasoning, resulting in predicted next 3D waypoints. A cross-space planning consistency loss is introduced to encourage directional agreement between the predicted map-space trajectory and expert action direction. Experiments demonstrate strong performance on OpenUAV and AerialVLN-S datasets with extensive ablations.
ASPIRE-VINS, an adaptive spline-based visual-inertial navigation system with robust 3D measurement residuals, combines adaptive knot placement, multi-resolution splines, and 3D measurement-space residuals to model motion-adaptively. It achieves competitive or lower trajectory errors than compared baselines in experiments under diverse motion and sensing conditions.
BrainWAM, a structured action-space coordination framework, combines semantic reasoning and predictive world modeling into two specialized pathways. It uses compact action representations to align these pathways and introduces an asynchronous rectified-flow inference strategy with decoupled video and action denoising. BrainWAM reaches state-of-the-art performance on both NAVSIM v1 and NAVSIM v2, outperforming VLA-only or WAM-only methods.
HumanoidVLN, a physics-grounded simulator and benchmark, was developed to address challenges in Vision-Language Navigation (VLN) for humanoid robots. The platform supports diverse humanoid embodiments and environments, and has been demonstrated on four robots with varying morphologies. The benchmark achieved the highest mean success rate of 43.55% and nDTW of 48.38 across four models and four embodiments.
AMR-Pose, an active LED marker-based relative pose estimation framework, was developed to improve robust vision-based relative localization in underwater environments. The framework combines Lie-group pose propagation, probabilistic marker association, and visibility-adaptive measurement fusion for accurate six-degree-of-freedom relative pose estimation. It adapts the estimation process according to marker visibility, maintaining geometric consistency and temporal stability during partial observations and visibility transitions.