Chinese regulators have blocked SoftBank's $5.4 billion acquisition of ABB Robotics. This action was due to concerns about the transfer of robotics data overseas for AI model training. China's significant presence in ABB's revenue and local manufacturing further influenced the decision. This event highlights growing geopolitical concerns in the global robotics industry. Furthermore, the text notes related actions by the U.S. government concerning foreign robot imports and technology.
A person signed up for a waitlist for a humanoid robot cleaning service in San Francisco, motivated by a desire for a futuristic experience. The robots currently operate with remote human piloting, using collected data to train future autonomous models. The service is expensive compared to the initial offer, and the robots are not yet fully autonomous for complex tasks. The experience revealed that while the robots made some impressive cleaning feats, they struggled with navigation and some tasks required human intervention. The company aims to eventually deploy these robots for various jobs, potentially improving conditions for those in the cleaning industry.
Agility Robotics began trading on the Nasdaq under the ticker AGLT, claiming to be the first pure-play legged humanoid company to go public. The stock experienced volatility, trading between a low of $8.40 and a high of $12.15, and closed slightly above its opening price at $11.49. Other robotics and technology stocks also experienced declines on the day. Agility Robotics has secured over $300 million in committed orders for its Digit 5 humanoids, with deliveries expected in 2027. The company is also exploring the addition of wheel-based robots to its product lineup.
Ambi Robotics helps customers prepare for peak season by providing support for increased logistics volume. They assist by offering training and implementing strategies to increase throughput without necessarily installing new systems. Ambi also reduces downtime by forward-deploying spare parts to customer locations. Customer trust is maintained through dedicated personnel for support, training, and maintenance. Furthermore, Ambi uses an internal AI system to assist employees in solving problems and onboarding new staff.
Embodied AI, a Swiss startup, unveiled AIR, a wheeled robot designed to work alongside factory staff without a safety cage. AIR features a torso with two compliant arms, designed to minimize harm during contact. The robot operates via remote teleoperation during initial training, and it is projected to achieve autonomy within a month. The company reports a 95% task success rate in paid pilots with manufacturers. Embodied AI plans to obtain CE marking in 2027 to comply with EU Machinery Regulation requirements.
Researchers at the University of Tokyo built a life-size robotic leg system using 3D-printed structures mimicking muscles, tendons, and ligaments. They utilized a single flexible TPU material, varying its geometry to achieve different mechanical properties for various body parts. Electric motors pull wires that actuate the legs, while the printed structures provide passive compliance. This method offers a shared design and fabrication process for compliant body parts. The work advances the design and manufacturing of robotic structures, though autonomous walking remains a future challenge.
Apple researchers trained a Unitree G1 humanoid using a system called HALO to navigate unseen rooms based on simulation-generated data. HALO successfully completed 91% of 360 single-task trials across six real-world environments without teleoperation. The system uses a multi-layered approach involving a vision-language model, a motion generator, and a motion tracker. Training involved generating 162,698 episodes using synthetic data from 1,000 procedurally generated rooms in Isaac Sim. Although successful in single tasks, chaining skills proved less reliable, highlighting ongoing challenges in real-world generalization.
Carnegie Mellon University introduced Pixel, a wheeled, two-armed robot with a build cost under $15,000, designed to learn from human demonstrations for coordinated movement. The project focuses on how co-designing the robot's body, controls, and data collection enhances learning. Pixel has demonstrated tasks like tidying and vacuuming, with a companion paper showing 87% average task progress. The system utilizes human-worn tools and sensor data to reconstruct scenes and translate movements, successfully handling complex interactions. The research aims to advance low-cost, accessible mobile manipulation platforms through human instruction.
NEURA Robotics is partnering with E.ON and enviaM to develop robots for electricity-grid operations, focusing on inspections and asset monitoring. The collaboration will involve training robotic skills using data and simulation at TUM RoboGym. The pilot program, with enviaM's MITNETZ STROM, is expected to begin in 2027. Initial applications will target inspecting equipment and monitoring critical assets in grid facilities.
Physicl raised $10 million in seed funding to apply its expertise in 3D rendering to improve robot training through more realistic simulations. The company aims to help model designers and robot builders overcome training difficulties by providing realistic synthetic data. Physicl's tool, Studio, allows users to create 3D scenes to generate data for training robots on specific, challenging tasks. The platform seeks to transform unstructured real-world information into high-fidelity physical representations for AI models. This approach allows AI to automatically generate necessary training data, addressing the challenge of manually curating data.
Two Xiaomi humanoid robots completed an eight-hour parts sorting trial at an electric vehicle factory. The robots successfully retrieved flexible plastic panels from bins and placed them into assembly line racks, demonstrating complex manipulation skills. The robots synchronized with the line's control system and corrected errors during the task. The trial involved variable presentation and flexible parts, unlike simple pick-and-place motions. This test served as a preliminary validation for one workstation, not a full fleet deployment.
Amazing Hand is an affordable, open-source robotic hand developed by Pollen Robotics. It features a compact design with motors inside the hand, eliminating external cables, and weighs 400 grams. The hand is constructed from 3D printed plastic and flexible shells, utilizing eight Feetech micro servos for its movement. It is positioned as an accessible experimental platform for hobbyists, education, and research rather than a premium dexterous hand. The project is available via ready-made kits or for independent assembly using printed components.
IHMC's Alex humanoid won the demonstration stage of the U.S. Army's xTechHumanoid competition at Marine Corps Base Quantico. The winning team, along with four others, advanced to structured evaluations with a chance to secure an Army prototype evaluation contract. Alex's success is attributed to hybrid autonomy, integrating AI decision-making with human operator control for complex missions. The competition aimed to assess humanoid capabilities for physically demanding tasks and reducing soldier exposure to danger. Future evaluations will focus on Alex's ability to navigate complex environments, which is a key design challenge.
The Allegro Hand V6 is a robotic hand from Wonik Robotics, designed for robotics research and humanoid development. It has a Humanoid.Guide Hand Score of 6 out of 10, reflecting a balance favoring dexterity over strength. Performance is assessed by separate ratings for Degrees of Freedom (4/5) and Strength (2/5), which describe different aspects of the hand's capability. The hand weighs 1.24 kg and is advertised at a price of USD 15,000. The product is in production and is significant for humanoid manipulation research.
Developed by MMLab at the University of Hong Kong, MM-Hand 1.0 is an open-source, modular research prototype with a lightweight design. It scores 5/10 overall, with a higher rating for degrees of freedom (4/5) than for strength (1/5). The design separates motors using tendons routed through flexible sheaths, and the hand weighs about 350 grams. The project is shared openly for research and adaptation, emphasizing its status as a research platform. Availability is as a prototype, and material costs are provided for budgeting context.
CHOHO Hand is a robotic hand manufactured by Choho Industrial in Qingdao, China. It achieved an overall Humanoid.Guide Hand Score of 8 out of 10, reflecting its state-of-the-art position. Its ratings are 4 out of 5 for Degrees of Freedom and 4 out of 5 for Strength, indicating balanced dexterity and grip performance. The product leverages Choho's industrial manufacturing background for its development. Currently listed at USD 5,000 and in production, it represents a commercially available entry in the robotic hand market.
Y4 is a stationary humanoid robot from Tlibot, featuring a humanlike torso and two arms designed for embodied AI data collection. Its capabilities focus on manipulation, allowing it to sort goods and assemble products with a 5 kg payload per arm. The robot operates using mains power and requires a customer-supplied computer with ROS 2 and API support for integration. It is supplied with joint modules and finger gripping claws for object handling. Y4 is currently listed as in production with an advertised price of USD 12,000.
Xiao Di is BYD's first self-developed humanoid robot, designed for customer-facing environments like car dealerships. It measures 161 cm tall and has 31 degrees of freedom with two five-fingered hands. Its capabilities include obstacle avoidance for navigation and sorting goods/assembling products for manipulation. The robot has a low Humanoid Guide Skill Score, with navigation rated 1/5 and manipulation rated 2/5. As a prototype, Xiao Di targets markets including car dealerships, education, and industrial uses.
Force is proposed as a natural command channel bridging motion imitation and object interaction for humanoid manipulation. Ctrl-F introduces a hierarchical framework that augments motion tracking with explicit 3-D force commands at designated contact points like hands or feet. This framework allows for the online adjustment of force commands based on object state and proprioception, which is difficult for force-blind trackers. Experimental results show Ctrl-F successfully handles tasks like dragging, pushing, and kicking boxes up to 20kg, adapting to changing loads. Consequently, the system enables effective locomotion manipulation requiring significant, sustained interaction forces.
MESA is an evaluation framework designed to precisely measure language-conditioned policy generalization in robotics. It offers five evaluation suites and four generalization axes: spatial configuration, object instance, object category, and subtask composition. The framework utilizes a pipeline called MESA-Gen for generating tasks and demonstrations at scale. Evaluation on seven baseline methods showed moderate performance, suggesting room for further progress. MESA rankings are broadly consistent with RoboArena scores, indicating relevance beyond simulation.
Robot imitation learning often suffers from limited action-labeled data despite abundant video demonstrations. ELA is an end-to-end framework that learns from video by optimizing latent actions jointly for policy learning and action decoding. It trains a latent action model, policy, and action decoder to leverage video data with few demonstrations. Experiments on manipulation tasks show ELA consistently surpasses previous latent action baselines. The method easily integrates with existing imitation learning approaches using lightweight decoders.
A trajectory planner is presented for humanoid robots to perform dynamic movements on complex, parameterizable terrains using a contact-implicit whole-body optimization framework. This method generates feasible motions by considering the robot's full dynamics and kinematics without requiring predefined contact sequences. A mathematical model approximates ground geometry with a smooth surface, modeling ground interaction through novel contact constraints for arbitrary surfaces. The approach is validated using nonlinear programming across various terrain topologies, with feasibility tested in simulation and on the ErgoCub robot. Dynamic trajectory generation is achieved by embedding high-level requirements into the optimization cost function.
Smooth trajectory planning is crucial for high motion performance and avoiding vibrations in robots, but existing smoothness requirements may be insufficient for soft continuum robots. The work introduces a $\mathcal{C}^\infty$-smooth trajectory generator to address this gap by creating infinitely differentiable paths. These trajectories ensure continuity in velocity, acceleration, jerk, and other derivatives between sequential states. The method specifically handles null boundary conditions where all derivatives at the start and end are zero. The approach was verified through simulations and hardware testing.
A framework called Proactive Human-in-the-Loop (PHIL) allows operators to intervene by previewing policy predictions before they cause mistakes. During intervention, PHIL uses an extended reality interface to display predicted trajectories and pairs rejected predictions with corrective actions. Preference-based post-training encourages the policy to favor these human corrections. Evaluation on manipulation tasks showed that PHIL increased success rates for diffusion policy and X-VLA compared to other methods. Policies corrected proactively also exhibited fewer unnecessary recovery behaviors in successful episodes.
OdorSpaces is a physics-based simulation platform and benchmark for visual-olfactory embodied navigation in 3D environments. It augments Habitat-Sim by simulating 3D odor transport and dispersion using the Fire Dynamics Simulator (FDS). The system allows agents to navigate photorealistic indoor spaces using both visual and olfactory inputs. Experiments focus on source-seeking tasks like gas-leak and fire localization. A multimodal reinforcement learning policy developed for this system outperforms baselines in visual-olfactory navigation.
OUCH is a wrist-worn device utilizing pneumatic actuation to deliver controlled, noxious mechanical stimuli for automated neurological assessment. It employs stacked pneumatic pouch actuators, an indenter, and a compliant spacer to ensure the indenter contacts the skin during actuation. Testing demonstrated that OUCH achieved significantly higher localized contact pressure compared to a commercial algometer while eliciting discomfort with less applied force. This work demonstrates a method for objective, uniform, and controlled haptic stimulation. The system provides a foundation for automated neurological assessment in hospitalized patients.
GRASP proposes a method for physics-based motion retargeting by combining sampling-based exploration with gradient-guided refinement. It extends tempered sequential Monte Carlo by fitting local quadratic cost models to archived rollouts to guide Markov chain Monte Carlo rejuvenation. The approach utilizes gradients from these rollouts to refine the sampling process without needing a differentiable simulator. GRASP-MPC scales this optimization to long demonstrations using warm-started receding-horizon planning. Experiments on hands and humanoids achieved accurate retargeting with minimal task-specific contact rewards.