The Toyota Research Institute (TRI) published a tutorial article introducing a statistical A/B testing method based on success/failure labels for robot policy evaluation.
Why it matters: The TRI authors systematically explain statistical pitfalls and reusable tools in robot policy A/B testing, with runnable Python examples.
Realman Intelligent announced it has completed nearly 500 million yuan in funding, jointly and strategically invested by industrial parties from multiple listed companies. The funds will go toward product R&D, construction of the AUTRON production line super factory, and global ecosystem development. The article states that its integrated joint modules exceeded an annual production capacity of 100,000 units in 2025, that its robotic arms passed CR L3 certification and achieved a mean time between failures of 50,000 hours, and that the GLN remote operation network and the Beijing Humanoid Robot Data Training Center have accumulated tens of millions of real-machine trajectory segments. The company serves more than 8,000 enterprise users worldwide.
Why it matters: The company describes its own progress on funding, production capacity, certification, and its remote operation network, giving readers a basis to understand how its hardware-plus-data strategy is actually advancing.
At the Beijing TV Spring Festival Gala, RealMan's RealBOT wheeled folding robot used remote teleoperation to open a refrigerator, take out soy sauce, pour soy sauce, and stir dumpling filling; it was the only company among the participating recording enterprises to complete on-site work via remote teleoperation. Its integrated joint module has an MTBF of 50,000 hours, opening the refrigerator door requires an instantaneous torque of about 14 kilograms, pouring soy sauce relies on visual guidance and high-frequency posture correction to achieve zero spills, and mixing the filling is controlled by a remote "isomorphic arm" that records data to feed back into training models for the remote operation network (GLN).
Linear Capital reposted LatePost's interview with Tashi Intelligence founder Chen Yilun, who laid out a technical path that differs from the mainstream: he rejects the VLA architecture that grows an action head on top of a VLM, and instead develops its own AWE (AI World Engine), aiming to represent physical quantities such as time, space, and force; it has also abandoned teleoperation data collection in favor of a self-developed collection device combining a wearable glove and a first-person-view camera.
Why it matters: The interview systematically presents technical judgments and a data approach that differ from the mainstream VLA path, giving readers a basis for understanding the specific grounds of the divergence in embodied AI approaches.
Skild AI proposes overcoming the robot data bottleneck by watching human videos, arguing that teleoperation is limited by environment diversity and real-time collection scale and cannot reach the data volume required by foundation models. Its approach is cross-embodiment learning: with less than 1 hour of robot data for fine-tuning, the model can learn new skills from video demonstrations alone, mapping human "grasping" onto the execution actions of different robot embodiments.
Realman is exhibiting at CES 2026 for the third consecutive year, presenting three foundational capabilities: hardware, data, and a remote operation network. It also completed a transoceanic real-time teleoperation demonstration from Beijing to Las Vegas, remotely controlling a RealBOT wheeled folding robot to perform tasks such as delivering items and handing over fruit.
Mentee Robotics demonstrated a MenteeBot learning to replace the battery in another MenteeBot without teleoperation, solely by observing a human demonstration. The learning process has three Real2Sim2Real stages: the robot records a human operation video from its own viewpoint, a foundation model reconstructs the task in simulation and then trains at scale using self-play and reinforcement learning, and finally its proprietary Sim2Real technology transfers the skill to autonomous execution on the real robot.
Zheng Suibing, founder and CEO of Realman, said on December 19 at the 2025 Corporate Social Responsibility Forum that the company is using a "remote workforce network" to let operators direct robots in real time to perform real tasks such as folding towels, utility tunnel inspections, and power control, while accumulating scenario data to feed embodied intelligence. Realman's integrated joint modules have annual production capacity exceeding 100,000 units, its robotic arms have passed CR L3 certification, and their mean time between failures has reached 50,000 hours. The company also co-founded the GPR open-source alliance in February 2023 and released a mobile composite robot development platform in 2024.
On December 19, Tashi Intelligence, founded just 10 months ago, held a livestreamed full-stack technology debut. The world's first robot to achieve autonomous embroidery was officially unveiled, completing thread pulling, needle lowering, and piercing motions with sub-millimeter precision to embroider the brand logo, while also demonstrating the autonomous implementation of flexible fine manipulation in wiring harness assembly.
Toyota Research Institute (TRI) reviewed eight years of polymer electrolyte research conducted by its AMDD project at MIT, producing about 40 papers and 7 doctoral dissertations and building a robotic experimentation platform that shortened tests that originally took about a month to completion within a day. The team also developed a multitask graph neural network to accelerate MD simulations and released the largest polymer simulation dataset, but the authors believe that AI can currently accelerate optimization within a class of materials, and that breakthrough polymers relying on new physical mechanisms still require human creativity.
Physical Intelligence research found that in generalization scenarios where only human demonstrations appear, fine-tuning π0.5 with human first-person videos mixed into robot data alone improved policy performance by about 2x, without using any explicit transfer learning mechanism.
Why it matters: Experiments show that merely mixing human videos into fine-tuning improves the policy by about 2x on scenarios with purely human demonstrations, and this capability emerges as pretraining robot data scale grows.
At IROS 2025, Dobot exhibited the world's first embodied intelligence robot platform integrating humanoid, robotic arm, multi-legged, and wheeled forms, using a brain-cerebellum dual-system architecture. Its humanoid robot passed high-temperature and heavy-load aging tests, with precision reaching ±0.05mm; the X-Trainer dual-arm teleoperation AI training platform shortens the training cycle for new tasks by 70%, to about 2 hours. Dobot also cooperated with Tencent's Tairos platform, and the multi-legged robot Hexplorer achieved fully autonomous operation with high-precision 3D mapping and dynamic obstacle-avoidance guided tours.
Sunday Robotics released the foundation model ACT-1, claiming it was trained without any robot manipulation data; instead, human motions were captured with a Skill Capture Glove and converted into robot data via Skill Transform with a 90% success rate.
At ROSCon Singapore, Intrinsic announced a partnership with Open Robotics to host the inaugural AI for Industry Challenge. Participants can use open-source tools and simulators to train AI models, develop solutions in Intrinsic Flowstate, and test them at real workstations at Intrinsic headquarters.
Tashi Intelligence has released World In Your Hands (WIYH), described as the world's first large-scale real-world embodied VLTA (vision-language-tactile-action) multimodal dataset, and plans to share it with the industry in December 2025.
Tianlian Robotics has published a video imitation learning training pipeline: extract reference motions with GVHMR/TRAM (or use motion capture datasets such as AMASS), then after human-to-robot joint mapping and training a motion orientation model, use PPO reinforcement learning in simulation with curriculum learning to progress from static poses to long-sequence complex motions, then optimize through domain randomization before deploying to the real robot, with fine-tuning via residual reinforcement learning or adaptive gains. The company says it has achieved key breakthroughs in dynamic force-control precision and stability, and that motions can also be reproduced in real robotic arm operation.
KEENON Robotics has officially released KOM 2.0 (KEENON Operator Model 2.0), which it calls the world's first VLA (vision-language-action) model for the service industry. The model uses a dual-system architecture with the K-Mind thinking system and the K-Act action system, is trained on hundreds of millions of real service interaction data points, and supports millisecond-level real-time decision-making.
LimX Dynamics has introduced LimX DreamActor, an embodied intelligence training paradigm that for the first time deeply integrates Real2Sim2Real, imitation learning, and real-robot reinforcement learning into a full pipeline of "real-world environment data collection—simulation data pretraining—real-robot data post-training." The system can complete multi-view data collection using only a consumer-grade camera; through 3D reconstruction and physical parameter alignment it builds a highly realistic simulation scene, then uses a small amount of real-robot reinforcement learning to complete cross-domain transfer, aiming to improve sample efficiency, convergence speed, and deployment stability.
Robot Lecture Hall reported on-site from the LimX Dynamics booth at the 2025 World Robot Conference. The full-size humanoid robot LimX Oli, 165 cm tall with 31 active degrees of freedom, made its first WRC appearance and went on sale, demonstrating outdoor live dancing and switching among multiple gaits such as a light quick step, model walk, and goose step, powered by a whole-body motion control algorithm combining traditional control, reinforcement learning, and imitation learning, with prices starting at 158,000 yuan.
Why it matters: The author attended WRC in person and breaks down the gait switching, disturbance-rejection recovery, and teleoperation details at the LimX Dynamics booth item by item, so readers can understand its full-chain technology roadmap.
SAIC Beijing and LimX Dynamics signed a strategic cooperation memorandum in Beijing, announcing the establishment of a joint embodied intelligence laboratory to conduct embodied robot algorithm R&D, system development, and toolchain deployment around automotive scenarios. LimX Dynamics will provide underlying technologies including robot bodies, whole-body motion control algorithms, VLA model frameworks, data engines, and toolchain platforms. The two parties will carry out humanoid robot data collection, training, and evaluation in automotive manufacturing scenarios, and gradually form a complete R&D process from data collection and model training to system testing.
TeknTrash Robotics has partnered with environmental services company Sharp Group to pilot the humanoid robot ALPHA (Automated Litter Processing Humanoid Assistant) at a recycling facility in Rainham, East London.
Figure AI released an end-to-end reinforcement learning walking controller that lets the Figure 02 humanoid robot learn a human-like gait in simulation by imitating human reference trajectories through rewards, including heel strike, toe push-off, and arm swing. Training runs in parallel across thousands of robots in GPU-accelerated physics simulation, collecting years of simulated data in hours, and combines domain randomization with kHz-level torque feedback control to achieve zero-shot transfer to the real robot; 10 Figure 02 units run the same neural network without individual tuning.
Why it matters: The original text provides details on the training method and sim-to-real transfer for Figure 02's reinforcement-learning-based human-like walking, offering insight into its technical approach.
Field AI Research Institute (FAIRI) is collaborating with NVIDIA to study robotic embodied question answering (EQ&A) in dynamic environments. The team uses the NVIDIA Omniverse platform, Isaac Sim, Omniverse Replicator, and the USD Search API to build physically accurate 3D office scenes, simulate changes in the environment over time, and plans to open-source this long-term embodied question answering dataset.
FieldAI claims to have developed the first risk-bounded Field Foundation Models (FFM) robot foundation model, which enables multi-form robots to operate autonomously in hazardous, off-road, and other unstructured industrial environments without prior maps, GPS, or cloud communication, and has already been deployed at hundreds of sites.