Researchers have developed a new imitation learning method, TGOD-SD, that significantly improves the ability of nursing robots to learn complex tasks from a single expert demonstration. The method addresses key limitations of traditional approaches, which often require large numbers of expert samples and struggle in dynamic, unstructured environments typical of healthcare settings.
Key takeaways
- TGOD-SD improves imitation success rates by an average of 64.6% compared to state-of-the-art methods.
- The method enables learning from a single expert demonstration, reducing dependence on large datasets.
- Trajectory quality improves, with a 32.61% average increase in correlation with expert demonstrations.
- Expected learning time is reduced to at least 62.5% of that of mainstream methods.
The proposed approach combines a guided diversity paradigm (TGOD) for generating diverse imitation trajectories around expert demonstrations, with a trajectory matching method based on Sinkhorn distance (SD) to select the best output. A sim-to-real transfer technique using joint angles allows deployment on physical nursing robots. The method is designed to work with reinforcement learning algorithms like SAC, using internal pseudo-rewards rather than manually engineered reward functions. This work offers a practical solution for automating care tasks in aging societies, potentially easing the burden on healthcare workers.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-01-13 · “基于引导多样性的护理机器人模仿学习”