TL;DR
This research presents a method for robots to learn core skills from human demonstrations that are "polluted" by obstacles. By applying an importance-weighting scheme that discounts movements made near clutter, the robot can infer the human's true intent—such as moving in a straight line—and generalize that skill effectively to new, obstacle-free environments.
The Challenge: Learning Intent from Polluted Demonstrations
Learning from Demonstration (LfD) allows robots to acquire complex tasks by mimicking humans. However, real-world demonstrations are rarely "clean." A human often has to navigate around clutter—like a drawer or a pile of tools—to show a robot how to perform a task.
Standard LfD algorithms struggle to distinguish between the essential components of a skill and the accidental detours caused by the environment. If a human reaches for a cup by curving their hand around a box, a robot might learn that "curving" is a fundamental part of the reaching skill. Consequently, when the box is removed, the robot continues to move in an inefficient curve. This paper aims to "filter out" these environmental influences so the robot learns the core skill, not the detours.
Proposed Method: Importance Weighted Skill Learning
The authors address this by building upon the CLAMP (Combined Learning from Demonstration And Motion Planning) framework. Their key innovation is Importance Weighted Skill Learning, which treats environmental constraints as a form of noise to be filtered out.
- Importance Weighting: The algorithm calculates a weight $w(x_i)$ based on the robot's proximity to obstacles. Movements within a defined "danger area" are given lower weights, while movements in open space are weighted more heavily.
- Probabilistic Inference: The robot uses these weights to estimate the parameters of its "trajectory prior" (the model of the skill). This can be done in batches or incrementally.
- Incremental Learning: To allow for continuous improvement, the authors developed a Bayesian approach using matrix-normal and inverse-Wishart distributions. This allows the robot to refine its skill model as new demonstrations arrive without needing to store all previous data.
- Task Reproduction: In a new environment, the robot performs "Maximum A Posteriori" (MAP) inference to find a trajectory that best fits the learned skill while remaining collision-free in the current scene.
Key Results and Insights
The researchers evaluated their approach using a 7-DOF JACO2 manipulator on reaching and placing tasks.
- Recovering the "Clean" Skill: In reaching experiments, the importance-weighted model successfully learned a mean trajectory that was significantly closer to a straight line than the unweighted model.
- Rapid Adaptation: The incremental learner could adapt to a "clean" environment after just three updates. Without importance weighting, the robot's "bad habits" (the detours) persisted much longer.
- Generalization: The robot successfully generalized skills to environments where obstacles were moved or removed entirely, proving it had learned the underlying task dynamics rather than just mimicking a specific path.
Limitations and Future Work
While effective, the system currently requires manual tuning of the "danger area" and assumes that all obstacle locations are known during the demonstration phase. Future work may focus on automatically learning these importance weights or extending the framework to handle more complex, non-linear skill dynamics.
Summary updated