Rhoda AI bets that predicting short sensory futures can organize robot behavior. A separate inverse dynamics model translates these predicted futures into actual movements. Recent reports show better video prediction leads to stronger bearing-unpacking performance, and tactile data extends this capability to touch. The core challenge for Rhoda AI is ensuring these training results translate to sustained factory operation. The company is focusing on translating sensory predictions into executable robot actions using video and tactile data.
HALO research investigates how humanoids can coordinate with human partners during physically coupled tasks like shared carrying. The project focuses on addressing the difficulty of independent learning undermining coordination by adjusting learning updates to align with a shared objective. Demonstrations show HALO collaborating with a human partner in tasks involving pushing, moving through spaces, and handling long objects. Comparative results indicate HALO performs well in confined-space transport and reduces post-halt drift relative to other methods. Despite promising results, the findings are based on limited trials, leaving broader reliability questions for real-world humanoid deployment.
AGIBOT introduced GO-3 Preview, an embodied foundation model that integrates reasoning, prediction, and action using the WRAM architecture. Learning is achieved through three stages involving Genie Evolver for post-training and Genie Master for deployment experience. The system uses various mechanisms to combine visual representations, future state prediction, and action generation to learn transferable robot capabilities. AGIBOT reports first-place benchmark rankings on RoboDojo, RoboTwin, and RoboColiseum and plans to fully open-source the model by the end of 2026.
AGIBOT's Physical Reasoning dataset connects planning with robot actions by showcasing how a single instruction requires varied physical execution based on context. Examples demonstrate that tasks involve decision-making regarding object quantities, memory for past searches, and assessing action preconditions like space. Annotations link high-level goals to detailed action trajectories and spatial relationships, providing multi-layered training signals for different reasoning abilities. The dataset includes unsuccessful attempts and levels to help evaluate a system's ability to handle deviations and recover. Ultimately, it aims to study how decision-making and physical manipulation integrate to enable more robust robot learning.
Bleu Robotics raised β¬5 million in pre-seed funding to develop software capable of teaching industrial humanoids new tasks in under an hour. The process involves gathering human demonstrations, followed by execution, with the claimed workflow taking about 40 minutes. The system utilizes "abductive learning" to minimize the necessary examples needed for task learning. The company aims to deploy this technology in industrial settings by offering software subscriptions to robot manufacturers. Future plans involve testing the system for reliability across changing production conditions.
The third edition of the Robotics and Simulation Devroom will be part of FOSDEM 2027. The event is scheduled for January 31, 2027, in Brussels, Belgium. Participation allows submissions for talks on open-source robotics and simulation topics, including ROS and Gazebo. Organizers encourage submissions of real-life robot demonstrations. The call for participation deadline is November 23rd.
AGIBOT introduced Worken, a framework to measure robots' productive work by focusing on autonomous performance rather than just deployment numbers. Worken calculates productivity using active robots, autonomous working hours, and an output efficiency ratio. The company also unveiled the AHA3 hardware platform, the GO-3 Preview foundation model, and the WITA-Omni 1.0 interaction model. They launched the G2 Max industrial robot and Link-U OS 1.0, emphasizing real-world deployment and integrated intelligence. These announcements focus on translating robotics innovation into measurable, practical work.
A Wiggle and Go system allows robots to estimate the physical properties of ropes and cables by observing their movement. This method enables robots to plan movements more reliably by predicting rope behavior without prior practice. Computer vision algorithms track the wiggling motion, which informs a model to estimate properties like length, weight, and stiffness. This capability promises to improve robot handling in fields such as manufacturing, construction, and search and rescue. The research team successfully trained their model in simulation before applying the learned estimations to real-world robotics.
A robot was developed to play the game Othello, stemming from an earlier, unsuccessful high school programming attempt. The robot features a Z-axis vacuum system to pick up clear tokens and place them on the board. It incorporates LEDs to show captures and flashing colors to indicate possible moves. The device is constructed using laser-cut components, 3D printed parts in PLA and TPU, and a Raspberry Pi camera for vision. This project demonstrates a sophisticated method for playing Othello.
Laborer M1 is a 187 cm tall, 86 kg bipedal humanoid from Wuhan, China, designed for industrial heavy load handling. It has a specified strength of 50 kg and 27 degrees of freedom with high torque capabilities. Its identified capabilities are obstacle avoidance for navigation and sorting goods through its arms. Despite its physical specifications, its skill score is low, indicating early-stage proficiency in navigation and manipulation. The robot is currently in production and costs USD 70,000.
Hobbs W1 is a wheeled humanoid robot from Noetix Robotics, standing 170 cm and weighing 75 kg, featuring a white body and a realistic silicone face. It has a maximum speed of 3 km/h and a six-hour runtime, equipped with 54 degrees of freedom. Its capabilities are focused on obstacle avoidance using LiDAR and sensors, and sorting goods, with limited manipulation skills. The robot is listed as in production with a price of USD 32,000 and is targeted for environments like service halls and reception areas.
E1 is a compact humanoid robot from Noetix Robotics, designed for research, education, and entertainment. It measures 136 cm tall and weighs 40 kg, excelling in physical mobility skills like obstacle avoidance, running, and stair climbing. Its strengths lie in locomotion, evidenced by a high navigation rating, while manipulation capabilities are more limited. E1 features five fingers on each hand and can sort goods, supported by advanced development hardware. Positioned as a platform for studying movement, E1 is currently priced at USD 21,000 and supports open-source development.
Bumi is a compact, consumer humanoid robot from Noetix Robotics, designed for home companionship, education, and experimentation. Weighing 17 kg and standing 98 cm tall, it features 21 degrees of freedom and basic physical capabilities like obstacle avoidance and standing after a fall. Its Humanoid Guide Skill Score of 3 out of 10 places it in the developing category, emphasizing mobility over advanced manipulation. Produced by a Beijing-based company, Bumi is priced at $1,400, aiming to provide affordable access to humanoid robotics for families and enthusiasts. The robot's capabilities focus on embodied robotics concepts like navigation and balance recovery rather than complex automation.
X Bot is a 167 cm tall, 60 kg bipedal humanoid developed by Yue Quan Bionics in China, featuring 51 degrees of freedom. It is equipped with electric motor modules and advanced computing from an NVIDIA Jetson Orin NX and an Intel NUC12 i7. The robot's primary demonstrated capabilities are obstacle avoidance and goods sorting, utilizing LiDAR and depth cameras for sensing. Despite its physical specifications, its overall skill score is low, placing it in the early stage category for navigation and manipulation. The platform is available for various applications, reflecting its status as a physical platform rather than a fully autonomous worker.
Gaia Hand 20 is a robotic hand from Stella Robot, a Chinese company, rated as capable with an overall Humanoid score of 6 out of 10. Its rating emphasizes dexterity, with a higher score for Degrees of Freedom (4 out of 5) than for Strength (2 out of 5). This balance is important for project assessment as dexterity does not automatically equate to grip strength. The hand is positioned as an affordable platform for embodied AI research. It is currently in production with an advertised price of USD 4,600.
Agility Robotics is opening a 50,000 square foot facility in Pittsburgh to centralize software, AI, deployment, and service teams for Digit projects. This expansion will also include an East Coast service center to complement their Salem, Oregon operations. The move supports the maintenance, software updates, and monitoring required for deployed humanoids, focusing on support capacity. The company has secured over $300 million in multi-year, conditional orders for Digit 5. This expansion adds 35,000 square feet to the existing Pittsburgh footprint.
Toyota and Waymo are advancing their autonomous driving collaboration to promote safer streets and increased freedom of movement. They will begin testing an autonomous Toyota bZ4X fleet in Japan starting in 2027, with plans for expansion to the UK and other Waymo markets. Toyota will manufacture these vehicles at its Takaoka plant, integrating Waymo's autonomous hardware. This partnership aims to scale safety-validated Level 4 autonomous vehicles globally by reducing human-error accidents. The collaboration also involves exploring future autonomous driving for personally owned Toyota vehicles and enhancing overall road safety.