Maxinsights records real-world tasks such as home kitchen work across 8 countries using its own capture hardware, labeling each clip with hand pose and language. The uniform setup keeps recording conditions fixed while environments vary, and the pipeline now serves more than 50 teams.
Imitation Learning
模仿学习
A method where agents learn strategies by observing expert behavior
Curated Intelligence
San Francisco startup Tau Robotics is offering a humanoid-robot cleaning service in which operators teleoperate the machines via head- and wrist-mounted cameras to wipe counters and fold laundry, while harvesting real-home data to train future autonomous models. The company prices cleanings below human rates to attract data-generating jobs and has completed only a few hundred sessions, underscoring how much human assistance home robots still require.
Verne Robotics released V0.5, an open-weight world action model built on the Wan2.2 video generator and trained on over 1.7M demonstrations, predicting the next second of arm, gripper and mobile-base motion. It learns new industrial tasks from under two hours of demos and degrades far less than Physical Intelligence's π0.5 under clutter and lighting shifts.
Chinese startup LiberAI released Liber-0, a robot foundation model trained on data from lightweight sensor gloves. It achieved a 36% success rate on new tasks without any robot demonstrations. Experiments showed that 10K hours of glove motion data outperformed 2M hours of internet video, significantly lowering the barrier for embodied AI data acquisition.
Reward AI demonstrates its OM-1 model's capability to maintain grip stability during humanoid locomotion. The model learns from human data via sensorized gloves without teleoperation or robot demos, enabling robust last-mile delivery tasks in complex environments like stairs.
Tesla is establishing an 'Optimus Academy' to train humanoid robots by deploying 10,000-20,000 units for autonomous practice and data accumulation. Elon Musk likens the training process to raising a child, noting that unlike autonomous driving, Optimus cannot rely on fleet data and must learn through observation and hands-on practice. Production has begun at the Fremont factory, with customer deliveries expected by late 2027.
Embodied AI data firm Maxinsights secured a $30M seed round and launched MaxGlove, a device capturing tactile and force data from human hands in real homes. This addresses the critical gap in robot manipulation training where vision-only data fails to capture force dynamics, aiming to build a million-hour dataset of high-quality operational signals.
A Purdue University team developed a robotic rehabilitation system combining generative AI and real-time force feedback to mimic human therapists and provide precise assistance to stroke patients. Currently in initial experimental stages, the system aims to extend rather than replace therapist capabilities, with plans for long-term clinical studies and potential industrial applications.
AI lab HUMXN launches a 'free home services for data' model, subsidizing plumbing and electrical repairs to record skilled workers' real-world tasks for training embodied AI robots. This approach addresses the limitation of internet text and images in capturing physical interaction details, initially rolling out in Minneapolis, Chicago, and Miami to explore compliant acquisition of high-quality real-scenario training data.
Amazon and Hugging Face demonstrate a closed-loop workflow for robotics using Strands Agents, LeRobot, and Storage Buckets. The system enables agents to record demonstrations, train policies directly from the Hub, and deploy them back to hardware, maintaining a consistent LeRobot data format throughout.