Shenzhen, September 9, 2026 鈥?Embodied intelligence company Liangyuan Xinchang has unveiled Light REACT, a whole-body resilient intelligence technology for humanoid robots, delivering the first milestone under its "mass deployment" paradigm.
Following the release of LightNav-0 on September 1, which marked the first achievement in "mass alignment," Light REACT tackles real-world whole-body control challenges by framing autonomous adaptation under faults and disturbances as an embodied In-Context Learning problem. The robot uses its recent whole-body interaction history as context to adapt to changing locomotion capabilities while model parameters remain unchanged. Without receiving fault labels or requiring manual switching of control modes, the robot adjusts its whole-body actions based on motion feedback, enabling autonomous movement even after partial joint failure 鈥?bridging the final gap toward large-scale deployment.
Autonomous Movement After Joint Failure
Mass deployment demands that robots adapt autonomously when operating conditions shift. When certain joints fail, identical motion commands can yield different outcomes, and existing movement patterns may become unsustainable. If anomalies require on-site intervention or model retraining, sustained robot operation becomes constrained. Light REACT integrates fault-adaptive walking, crawling, and fall recovery into a single control policy: when partial joints fail, the robot coordinates remaining joints to adjust gait and power output; when standing becomes impossible, it transitions to crawling using arms and other body parts; after falls caused by external disturbances, it attempts recovery based on its current capabilities.
In internal simulation tests involving complete power loss in one leg, Light REACT demonstrated the ability to continue walking by adjusting whole-body motions, or resort to crawling, all without manual mode switching. From gait adjustments to full-body behavioral reorganization, this whole-body resilience enables robots to leverage remaining capabilities for autonomous movement after failures.
Embodied In-Context Learning: Adapting to Faults from Interaction History
When a joint fails, the change manifests in motion feedback: expected actions go uncompleted, and actual movements diverge from the original intentions. These discrepancies are embedded in the recent whole-body interaction history, providing the policy with evidence to assess current locomotion capabilities. Light REACT does not rely on labels indicating fault location or severity; instead, it uses this history to infer which actions remain viable and adjusts subsequent control accordingly. As new interactions enter the context, the policy continuously adapts based on real-time feedback.
With model parameters held constant and interaction context continuously updated, robot actions adjust in tandem 鈥?this is the embodied In-Context Learning problem at the heart of Light REACT, and the core mechanism driving robot adaptation during deployment. Liangyuan Xinchang names this mechanism "Whole-Body Context Learning," extending context-based adaptation across whole-body control, from stride length and force adjustments during walking to behavioral transitions between walking, standing up, and crawling. Operational experience flows directly into the policy context, serving as the basis for real-time responses, eliminating the need for on-site retraining or manually designated response modes.
Learning to Use Context in Training, Adapting Autonomously Through Interaction in Deployment
This capability stems from deliberate design during the training phase. The Liangyuan Xinchang team synthesized extensive whole-body interaction context data in simulation, covering walking, crawling, and fall recovery under diverse fault conditions. Using this data, the team trained Transformer policies to leverage continuous interaction history, infer current locomotion capabilities from motion feedback, and adjust whole-body actions accordingly. During deployment, robots do not require fault labels or parameter updates; they rely on recent whole-body interaction history as context to autonomously adapt to changing capabilities. Learning to utilize context during training, and continuously adapting through interaction during deployment, forms the foundation of Light REACT's whole-body resilient intelligence. The framework's name derives from its research focus: REsilient humAnoid ConTrol.
Making Autonomous Adaptation a Foundational Capability for Mass Deployment
As robots move toward mass deployment, each unit encounters different disturbances, faults, and operating conditions. A single model must be able to leverage each robot's unique interaction experiences to address these variations. Light REACT brings this adaptive capability into whole-body control, enabling robots to adjust behavior based on their own motion feedback. For Liangyuan Xinchang, mass deployment should allow autonomous capabilities to scale alongside fleet size, reducing reliance on on-site personnel for every anomaly. Experience accumulated through continuous operation will also feed back into subsequent training and model iteration, creating a data flywheel.
The three-stage paradigm collectively aims to ensure capabilities acquired during training continue to function effectively in the real world, while improving through practical application. Light REACT represents the first achievement in this direction for mass deployment: making autonomous adaptation a fundamental capability after robots are deployed, allowing real-world operation to continually drive the next round of capability enhancement.