Residential concerns arose when Amazon's delivery drones, specifically the large MK30 models, generated bothersome noise and flew lower than legally permitted. Residents experienced significant annoyance from the high-pitched whine of the rotors, which acoustic research indicates is more irritating than car or airplane noise. Delivery drone companies are expanding rapidly, and while they promote positive uses, communities fear corporate overreach and safety issues. Activists and local governments have responded by challenging regulations and advocating for community engagement regarding drone operations. Ongoing disputes remain concerning noise, safety, and the legal regulation of airspace above private property.
GlobalLogic and Haystack Robotics developed an autonomous mobile robot (AMR) to provide hospital-grade UV disinfection with minimal human intervention. This robot utilizes real-time perception and edge-based intelligence to navigate complex hospital environments safely. The platform was designed to be an expandable foundation, allowing for future applications beyond disinfection. The robot features two modes: autonomous navigation and a "follow me" mode for caregiver assistance. Challenges for hospital robotics include ensuring a justifiable return on investment, meeting medical compliance standards, and navigating crowded human workflows.
LiberAI introduced Liber-0, a world action model pretrained on over 2.2 million hours of human experience to enable zero-shot robot manipulation. This model combines human video, glove-recorded motion, and embodiment alignment to transfer human manipulation skills to robots. In a zero-shot evaluation of 60 GPT-6 tasks, Liber-0 achieved 36.0% overall success without task-specific robot demonstrations. The system achieves this by grounding human experience in measured motion and aligning it with the robot's body representation. Furthermore, incorporating glove recordings can reduce the need for specific robot demonstrations, potentially saving significant data collection efforts.
NYU released the design for MABEL, an open-source, mobile bimanual robot with 56 degrees of freedom. The robot features two 7-DOF arms, two 17-DOF hands, and a pitching torso, with a total open-source parts list cost of $8,056. MABEL is controllable via the Apple Vision Pro and an ACT policy trained on demonstrations. The system is released under open licenses, including CAD, firmware, and simulation models. The project aims to advance humanoid robotics by combining low-cost components into a complex, teleoperated platform.
RoboColosseum is an open initiative providing access to over 100 robots and 10,000 free evaluation hours for testing robot AI on physical hardware. Unitree supports this project, highlighting its G1 platform for humanoid robotics. Initial tests revealed a significant gap between partial progress and completing full humanoid tasks, with models showing only partial success on tasks like carrying objects. The initiative aims to standardize testing by using common protocols and releasing evaluation recordings for inspection. Researchers can apply for access to this infrastructure to test model performance in varied real-world settings.
KIST's KAPEX humanoid reportedly utilizes over 90% Korean-made core components. This development is part of a collaboration with LG and other partners for commercialization. Various Korean entities are involved in different aspects of the project, including platform, AI, batteries, and manufacturing. KIST has transferred core technology to LG Electronics, with roles assigned across the partnership. The goal is to develop a domestic humanoid industry by demonstrating reliable system performance.
A new system, UQ-DAAAM, corrects unreliable object descriptions stored in a robot's memory by incorporating uncertainty quantification. It achieves this by comparing multiple viewpoints and captions of the same object to assess the robot's confidence in its descriptions. The system refines captions by selecting additional views that reduce uncertainty, leading to more accurate memory. This process boosts confidence and improves the accuracy of object descriptions by up to five percent. Ultimately, this method enhances a robot's ability to perform complex tasks by ensuring it remembers what it has observed with greater reliability.
Pudu Robotics unveiled its "One Brain, Multiple Embodiments" strategy and full product portfolio at the European Partner Summit in Amsterdam. This strategy aims to build a shared technological foundation across various robot forms, combining reusable AI with purpose-built hardware. The portfolio spans six product lines, including service delivery, commercial cleaning, and home lawn care. Core technologies like PuduFM and PuduAgent OS enable shared AI capabilities across different robotic platforms. Pudu is focusing on expanding robotics into diverse real-world applications across Europe.
AGIBOT unveiled GO-3 Preview, an embodied foundation model that supports learning across the entire lifecycle of pre-training, post-training, and post-deployment. The model achieved first-place rankings on the RoboDojo, RoboTwin, and RoboColiseum benchmarks. GO-3 Preview unifies task understanding, future-state prediction, and action generation through its embodied-native WRAM architecture. Post-training is facilitated by Genie Evolver, which uses physical robots, simulation, and world models for continuous refinement. The system incorporates Genie Master to enable robots to learn from real-world operational experience and improve their action policies autonomously.
Agility Robotics completed its SPAC merger and will trade on Nasdaq under the ticker AGLT. The company manufactures Digit, a bipedal humanoid already deployed in warehouse settings for clients like Amazon and Toyota. Digit 5 is in development to incorporate safety features allowing it to work near people, with initial deliveries targeted for the first half of 2027. Agility has tentative orders totaling $300 million from prospective customers, though revenue remains small relative to spending. The company employs about 500 people and is expanding its manufacturing and engineering operations.
Revo 1 is BrainCo's bionic hand for robotics, developed using their prosthetic hand technology. It features a five-finger design, weighs 0.54 kg, and is constructed from aluminum alloy and food-grade plastic. The hand incorporates six precision micro motors and receives an overall Hand Score of 5 out of 10, placing it in the developing category. It is currently in production with a listed price of USD 9,500.
The RY H6 is a compact, lightweight robotic hand designed for smaller humanoid platforms, featuring five fingers weighing 278 grams and measuring 155 millimeters. It utilizes brushless coreless motors and a force and position hybrid control system for operation. The hand is rated modestly in terms of dexterity (2 out of 5) and strength (1 out of 5), reflecting its focus on a small physical envelope. It is currently listed as in production with an advertised price of 1,300 USD. The product offers a commercially available option for teams developing smaller robotic systems.
The Ruiyan Robotics RY H5 is a robotic hand assessed with an overall score of 4 out of 10. Its evaluation includes a Degrees of Freedom rating of 2 out of 5 and a Strength rating of 2 out of 5, with both contributing equally to the total score. The hand has an all-metal body and weighs 0.8 kilograms, manufactured in China. The assessment indicates the product is a developing item rather than a leading dexterous hand. The RY H5 is listed in production with an advertised price of USD 2,500.
Unitree Robotics' Dex3-1 is a three-finger robotic hand designed specifically for the G1 humanoid platform. It features a compact design measuring 175 × 88 × 77 millimeters and weighing 0.71 kilograms. The hand incorporates miniature brushless motors, and its overall Hand Score is 3 out of 10, placing it in the developing category. Research should focus on task performance rather than appearance, evaluating dexterity and strength alongside integration requirements. The hand is currently in production with an advertised price of $1,600 USD.
The research introduces Jev-AD, which integrates Jev, a typed decision model, with a language world model for safe autonomous driving. The language world model predicts the movement of nearby vehicles over the next two seconds to generate foreseen consequences for candidate actions. This foresight allows Jev to compare potential outcomes directly rather than implicitly inferring consequences from coordinates. In simulations, this foresight significantly reduced collision rates and improved the model's ability to achieve passenger intent. Real-world testing further demonstrated that this approach lowered collision rates in actual driving scenarios.
Force is proposed as a natural command channel between motion imitation and object interaction, specifying both contact location and force magnitude. Ctrl-F is a hierarchical framework that augments motion tracking with an explicit 3D force command at contact points like hands or feet. This framework includes a unified controller and a high-level force planner that issues online force commands based on object state. In experiments, Ctrl-F successfully enabled humanoid robots to manipulate objects up to 20kg through dragging, pushing, and kicking. This system demonstrated the ability to adapt force commands online as external loads changed during manipulation tasks.
MESA is an evaluation framework designed to precisely measure language-conditioned policy generalization in robotics. It introduces five evaluation suites and four generalization axes: spatial configuration, object instance, object category, and subtask composition. The framework uses MESA-Gen to generate tasks and demonstrations for large-scale evaluation. Testing seven baseline methods, the strongest model achieved 52% on MESA and 49% on the BiMESA benchmark. MESA rankings align with RoboArena scores, suggesting relevance beyond simulation.
ELA is an end-to-end framework designed to learn latent actions from abundant video demonstrations for robot imitation learning. It addresses limitations in prior latent-action methods by jointly optimizing a latent action model, policy, and action decoder. This approach allows policies to effectively utilize video data with minimal action-labeled demonstrations. Experiments on manipulation tasks demonstrate that ELA consistently surpasses previous latent action baselines. The framework integrates easily with existing imitation learning methods through lightweight decoders.
A trajectory planner is presented for humanoid robots to achieve dynamic motion on complex, parameterizable terrains using a contact-implicit whole-body optimization framework. This planner generates feasible movements by considering the robot's full dynamics and kinematics without requiring predefined contact sequences. A mathematical model approximates ground geometry as a smooth surface, incorporating novel contact constraints for arbitrary smooth interactions with the ground. Nonlinear programming is employed to plan motions across various terrain topologies, and high-level requirements are integrated into the cost function for dynamic trajectory generation. Feasibility tests were conducted on non-coplanar terrains in both simulation and on the ErgoCub robot.
Smooth trajectory planning is crucial for high motion performance and vibration avoidance in robotics. The required smoothness level depends on the robot's degrees of freedom and system order, and this requirement may be higher for soft continuum robots than previously assumed. The paper introduces a $\mathcal{C}^\infty$-smooth trajectory generator to produce infinitely differentiable paths while respecting kinematic limits. This method generates trajectories with continuous velocity, acceleration, and higher-order derivatives. The approach is verified through simulations and hardware testing, including the case of null boundary conditions.
A framework called Proactive Human-in-the-Loop (PHIL) allows operators to intervene by previewing policy predictions and displaying trajectories in an XR interface. During intervention, rejected predictions are paired with corrective actions, and preference-based post-training steers the policy to favor these corrections. This method was evaluated on four real-world manipulation tasks using diffusion and vision-language-action policies. PHIL significantly increased success rates and demonstration budgets compared to existing methods like HG-DAgger. Policies trained with proactive corrections also exhibited fewer unnecessary recovery behaviors.
OdorSpaces is a physics-based simulation platform and benchmark for visual-olfactory embodied navigation in 3D environments. It integrates 3D odor transport and dispersion, simulated by the Fire Dynamics Simulator, into Habitat-Sim for navigation. The platform uses both visual and olfactory observations to guide agents through photorealistic indoor settings. It is benchmarked on tasks like locating invisible gas leaks and early-stage fires. A multimodal reinforcement learning policy developed for this system significantly outperforms existing baselines.
OUCH presents a wrist-worn device using pneumatic actuation to deliver controlled noxious mechanical stimuli for automated neurological assessment. It achieves objective, uniform, and controlled stimulation by combining pneumatic pouches, an indenter, and a compliant spacer. Experimental results showed the indenter concentrated load significantly more effectively than a commercial algometer tip. Preliminary testing indicated that the device elicited discomfort with approximately 89% less applied force than the algometer. This work establishes a foundation for objective neurological assessment through wearable haptic stimulation.
GRASP proposes a method for physics-based motion retargeting of human demonstrations under contact dynamics. It combines sampling-based exploration with gradient-guided refinement, avoiding the need for a differentiable simulator. GRASP-TO extends tempered sequential Monte Carlo by using gradients from local quadratic cost models to guide MCMC rejuvenation. This approach scales to long demonstrations via warm-started receding-horizon planning. Experiments show accurate retargeting for dexterous hands and humanoids, achieving 92.8% success on bimanual demonstrations.