WCM robot model hits 73.8% success with human-in-the-loop teaching
Robots can now learn from users in real-time, achieving 73.8% task success rates.
Current robot control systems lack transparency and user adaptability — they execute instructions but don't explain why actions are chosen or allow easy correction. To solve this, Yuzhen Chen and KC Zhou developed the World-Cognition Model (WCM), built on the SLAK architecture (Sensing, Logic, Action, Knowledge). WCM separates perception, reasoning, control, and memory, and runs them asynchronously, enabling dialogue and task execution to happen simultaneously without blocking. A key innovation is the human-in-the-loop teaching mode, where users can guide the robot through difficult or multi-step tasks. These teaching episodes are refined into chain-of-thought supervision, continuously improving the model over time.
In real-world trials across nine human-robot interaction tasks, WCM achieved a 73.8% average success rate. This includes tasks that were held out from chain-of-thought fine-tuning and a long-horizon task that the robot learned purely through user teaching. The work points toward more intuitive, collaborative robots that can be trained on the fly rather than requiring static programming. For professionals, this means robots that adapt to dynamic environments and user expertise, reducing deployment friction in manufacturing, healthcare, and service industries.
- Based on SLAK architecture separating Sensing, Logic, Action, and Knowledge with an asynchronous runtime for concurrent processing
- Achieves 73.8% average success rate across nine real-world human-robot interaction tasks, including held-out and long-horizon tasks
- Human-in-the-loop teaching mode lets users interactively correct and teach, with episodes distilled into chain-of-thought supervision
Why It Matters
Interactive teaching makes robots adaptable in dynamic environments, reducing need for constant reprogramming in industry.