Wetour Robotics Unveils Orchestra System for Enhanced Robot Learning

Wetour Robotics Limited, a company based in Austin, Texas, has unveiled a new development in physical AI training with the Orchestra system, which combines surface electromyography (sEMG) with first-person vision to capture richer human-hand data for robot learning. The system addresses limitations in vision-only capture, such as the inability to observe the force applied by a hand and the loss of hand data when the hand is hidden behind an object or out of the camera's view.

According to Nan Zheng, CEO of Wetour Robotics, "Physical AI needs more than just video. A camera can show where a hand moved, but not how hard it worked, and it can go blind at the exact moment contact happens. Orchestra is designed to add force-related information and continuity to human demonstration data."

Orchestra integrates Conductor, an 8-channel sEMG wristband, with VisionLink, a first-person camera. Vision provides spatial position and scene context, while sEMG captures muscle activation, remaining available during visual occlusion and potentially providing a physical signal before visible movement begins. The system is designed to fuse both streams locally into a synchronized record of movement, effort, and action timing.

In a test involving an internal carrying task, the vision-only pipeline failed to locate the hand in 21.8% of frames, including a longest continuous dropout of 4.32 seconds. The architecture is designed to use sEMG as a complementary signal during these missing visual intervals, with the cross-modal correction strategy currently in validation.

Key features of the demonstration include: - Continuous digital hand: The current on-device model outputs 20 joint angles rather than a small set of preset gestures. Processing a one-second sEMG window requires 50.4 milliseconds with zero lookahead, reflecting model processing time, not end-to-end system latency. - Force as a measurable dimension: The force-estimation pipeline converts sEMG into scale-calibrated, kilogram-equivalent grasp-force estimates. Calibration uses a scale as the reference and fits the response to the wearer and wearing session. The pipeline has been validated end-to-end on synthetic data; validation with live wristband force recordings remains under development. - Action-ready data: The system is being developed to layer events such as reach, grasp, hold, and release over continuous hand motion, creating structured demonstrations for imitation learning and robot training.

Five real-world tasks were demonstrated, including packing a lunch box, sorting pills, disassembling a pen, measuring a drone with calipers, and installing a drone propeller. These tasks combine precision, changing hand effort, and frequent visual occlusion.

The company believes that adding force-related data and greater resilience to visual occlusion can make human demonstrations more useful for fine manipulation, compliant control, and future human-robot collaboration in real environments.

Demonstration videos are available on Wetour Robotics’ website and social media channels under Wetour Robotics and @WETO_IR_TEAM.

Wetour Robotics Limited is developing Orchestra, a Physical AI platform that connects intelligent agents to the physical world through wearable sensors, visual intelligence, edge AI computing, and connected machines. Conductor is the company’s wrist-worn neuromuscular interface, and VisionLink is its visual intelligence module.

The press release also includes forward-looking statements regarding the development, integration, validation, performance, applications, and commercialization of Orchestra, Conductor, VisionLink, sEMG-vision fusion, force estimation, multimodal data capture, and humanoid or embodied-AI applications. These statements are subject to risks including technology development and integration risk, differences between synthetic and real-world data, calibration and accuracy limitations, competition, customer acceptance, capital requirements, and other risks.