The ARM Institute's member consortium has expanded to include over 500 organizations nationwide. This consortium, along with the Institute, collaborates with the U.S. government to promote robotics and AI leadership in U.S. manufacturing. Members have collaborated on numerous technology projects and initiatives, including the RoboticsCareer.org. The Institute focuses on developing innovative robotics and physical AI solutions and preparing the U.S. workforce for advanced manufacturing careers. Membership provides members with access to funding, events, and a digital community.
PrimeBOT's 10,000th personal robot was completed on November 8, 14 days after the 5,000th unit was finished. This rapid production increase resulted in a daily capacity of 240 units. The company emphasizes that this milestone reflects manufacturing speed rather than sales figures, as no sales data has been published. PrimeBOT robots are designed for companionship and learning, featuring customizable and updatable personalities. User data suggests a high level of engagement with the robots once they are in use.
Researchers developed FLEX, an AI-assisted learning platform to enable robots to reason beyond literal commands from natural language. FLEX uses a large language model (LLM) to draw larger generalizations from human instructions, improving motion-planning. The process involves three steps: Add, Generalize using the LLM, and Mask to update the robot's understanding of the environment. In simulations, the robot successfully inferred higher-level intentions, such as avoiding electronics when told not to spill coffee. This method allows robots to adapt to unspoken preferences, making them more intuitive for human interaction.
The poster is developing an offline monitor for autonomous mobile robot (AMR) deployment traces. In simulations, the poster observed changes in multi-sensor state signals preceding a modeled failure boundary. The core question posed is whether similar precursors to docking or localization failures are observable in real-world AMR fleets. The poster is soliciting information regarding specific signals to monitor before docking misses or localization resets occur. This post seeks real-world observations related to telemetry data preceding AMR failures.
Optical motion capture provides robots with an independent reference for training and hardware validation, addressing the measurement problem in imitation tasks. It allows researchers to distinguish between a policy's quality and the system's physical execution by establishing an external truth. Different capture methods, such as optical, inertial, and markerless, offer varying trade-offs in accuracy, latency, and handling of occlusions. Valid evaluation requires comparing the reference trajectory with the robot's estimated state to measure divergences like foot-placement errors or trajectory deviations. Ultimately, precision tracking is valuable for separating demonstration collection from simulation-to-real testing and for verifying autonomous navigation.
StarBot Robotics has opened reservations for the Spark, a three-foot humanoid, starting at $7,999 plus a $200 deposit. The company is offering the first 100 confirmed orders a starting price of $5,999. Spark features 22 degrees of freedom and a 24-volt battery pack providing up to three hours of runtime. The platform is positioned as an open development tool for education, programming, and research. StarBot claims the Spark is the first US-made developer humanoid in its size and price range.
Y Hand M2 is a bionic robotic hand developed by Yuequan Bionics in Beijing, China. It scores 9 out of 10 overall, with a top rating for Degrees of Freedom (5/5) and a strong rating for Strength (4/5). The design uses a structure inspired by human anatomy, combining rigid and soft components. Actuation is achieved through magnetic electric drive artificial muscles integrated into compact units. The hand measures 200x90x40 mm, weighs 0.7 kg, and features five fingers, and is currently listed for production at USD 15,000.
Six Agibot G2 wheeled humanoid robots completed 1,730 tasks during a 24-hour trial at a Shanghai convenience store. The robots successfully handled various tasks, including customer service interactions, product pickups, and shelf restocking. However, only four machines passed the final assessment for the trial. Specifically, the robots assigned to cleaning and shelf inspection failed to meet performance criteria. The experiment provided a comprehensive view of retail capability across changing customer activities and store conditions.
Neo1 is a 400 cm tall, 1,200 kg industrial humanoid developed by Beijing's Neodra Dynamics. It features a system combining hydraulic and electric actuation with 41 degrees of freedom and a 500 kg combined arm payload. The robot is designed for industrial applications, supporting functions like obstacle avoidance, sorting goods, and product assembly. Its capability score is low, reflecting developing performance in navigation and manipulation tasks. Currently listed as a prototype, Neo1 targets markets such as heavy manufacturing and mining.
CRAFT Hand is an open-source robotic hand developed by researchers at the University of Illinois Urbana Champaign and UC Irvine. It features a hybrid architecture using rigid links and soft joint material actuated by 15 Dynamixel servo motors and rolling contact joints. The hand measures 200 × 95 × 40 mm and weighs 0.6 kg, utilizing 3D printed rigid links and compliant joint material. It scores 4 out of 10 overall, with a rating of 3/5 for degrees of freedom and 1/5 for strength. The project is shared as open source, providing documentation for research replication.
Handroid is a desktop robotic hand developed by researchers at UNC Chapel Hill and Stanford University. It features a 3D printed plastic structure with modular actuators designed for dexterity and adaptable mechanical design. Assessment scores show a high rating for degrees of freedom (5/5) but a lower rating for strength (1/5). The hardware includes five fingers, joint position feedback, and an integrated control board, weighing 2.05 kilograms. The project emphasizes modular construction and provides design files for research reproducibility.
RP1 is a 130 cm tall, 34 kg bipedal humanoid developed by Shanghai RoboParty Technology Co., Ltd. It functions as an open, modifiable platform for robotics research and embodied AI. Physical capabilities include obstacle avoidance, standing up after a fall, and sorting goods, utilizing 30 degrees of freedom and 10 fingers. The robot operates on a software foundation of Linux, ROS 2, and RoboParty PartyOS. It is positioned as a prototype with a low skill score, emphasizing its role as a development platform.
Human interventions offer valuable feedback for improving robot behavior, but current methods rely on inaccurate human models. The research found that humans intervene when agent progress stalls, and their corrections are often suboptimal. Q-OIL introduces a progress-based model of human intervention to create realistic benchmarking in simulation. This method uses optimistic bonuses to guide exploration toward human corrections, learning more efficiently than competitive baselines. Q-OIL achieved 10-50% higher success rates in simulation and on real robots for bimanual manipulation tasks.
Meituan-Robotics-0 is a vision-language-action foundation model designed for stationary-base desktop manipulation. It integrates 147 million vision-language samples and 15,060 hours of robot trajectories across twelve embodiments. The model uses a three-stage training recipe involving action-token pre-training and flow-matching to align multimodal representations with robot behavior. Rollback is introduced to provide supervision for policy training without value learning. The system includes features for heterogeneous end-effector interfaces and event-driven task management for improved execution.
Flow-matching policy sampling aims to enforce constraints on deployed policies at inference time without discarding learned behaviors. The method combines rejection, which ensures plausibility, and projection, which ensures efficiency. CS-Pi introduces a "draw-project-renoise" sampler that projects draws to the nearest safe action, adds scaled noise, and re-runs the policy for safety. This approach successfully balances the efficiency of projection with the plausibility of rejection. Experiments show CS-Pi achieves high safe success rates across various robotic tasks.
ForPC is a force-aware whole-body control framework designed to achieve precise humanoid cart pushing. It uses two-stage reinforcement learning to regulate cart velocity by inferring force and state latents from onboard observations without external sensors. The policy coordinates arm, waist, and leg adjustments with base-velocity commands, penalized for internal opposition and grasp slip. Simulation results show ForPC reduces velocity and cross-track RMSE significantly compared to locomotion-only baselines. The trained policy successfully transferred to a physical robot to push a cart with a 30 kg payload.
Variation in state regulation of automated vehicle deployment in the U.S. stems from the lack of a unified federal certification framework. The paper compares how U.S. states regulate SAE Level 4-5 automated driving systems against international practices. Three state certification models were identified: baseline compliance, administrative self-certification, and permit-based authorization. International trends favor scenario-based testing, safety-case documentation, and lifecycle oversight. The research develops a framework to compare state approaches based on regulatory involvement and governance complexity.
VICTR is a framework designed for in-context learning in Vision-Language-Action (VLA) models through progress-informed context retrieval. It introduces an In-Context Value Function Estimator (IC-VFE) to estimate task-relative progress based on a reference demonstration, language, and current observation. This progress estimation guides the selection of visually relevant demonstration chunks for conditioning the action generation policy. Applying VICTR significantly improves performance on unseen out-of-domain real-world manipulation tasks. Specifically, VICTR increased the success rate from 6.2% to 23.8% and more than doubled the task progress rate from 28.9% to 55.5%.
A multi-agent object-centric reinforcement learning framework is presented for geometric assembly tasks. Each part is treated as an agent learning a shared policy for continuous assembly control. A graph neural network models the pieces and their relationships to predict translational and angular velocities. Proximal policy optimization with a centralized graph critic learns motion and temporal coordination across the agents. The method demonstrates concurrent assembly in simulation and successful transfer to real robot execution.
STRIDE learns humanoid locomotion by using acceleration-based rewards to train robots with only onboard Inertial Measurement Unit (IMU) data. Instead of tracking base velocity, the method rewards the robot for following a commanded acceleration profile. The approach utilizes the off-policy RL algorithm XQC combined with high gain joint position controllers and left-right symmetry augmentation. In simulation, STRIDE successfully taught robots to walk on Unitree G1 and Booster T1 humanoids within fourteen minutes. This method significantly reduces training falls by 81% compared to other approaches.
Millimeter-wave radar offers an alternative for human state sensing in challenging environments, as it operates without ambient light and penetrates smoke. Limitations arise from compact arrays, which result in joint-localization errors of about decimeters at several meters. A new framework models the human and environment as a time-varying scattering field to establish a Cramer-Rao error floor for localization. Experiments show that techniques like frame fusion and multiple viewpoints can reduce localization errors by moving them in predictable directions. This provides a basis for robotics to decide between observation design and estimator capacity for reliable mmWave sensing.