Towards Robust Skill Generalization: Unifying Learning from Demonstration and Motion Planning

Muhammad Asif Rana, Mustafa Mukadam, Seyed Reza Ahmadzadeh, Sonia Chernova, Byron BootsCoRL 2017 · OpenReview

Abstract: In this paper, we present Combined Learning from demonstration And Motion Planning (CLAMP) as an efficient approach to skill learning and generalizable skill reproduction. CLAMP combines the strengths of Learning from Demonstration (LfD) in capturing complex motion styles with the robustness of Motion Planning (MP) in handling environmental constraints. By formulating skill reproduction as a probabilistic inference problem, we enable robots to adapt learned behaviors to novel scenarios and obstacles.

1. Introduction

Traditional robotics often separates motion generation into two distinct paradigms: Motion Planning (MP) and Learning from Demonstration (LfD). MP excels at obstacle avoidance and satisfying geometric constraints but struggles with tasks that are difficult to define mathematically. Conversely, LfD allows robots to acquire complex skills from human guidance but often lacks the flexibility to generalize to new environments or avoid unexpected obstacles. This paper introduces CLAMP, a unified framework that bridges this gap by treating skill reproduction as a Maximum A Posteriori (MAP) inference problem.

2. The CLAMP Framework

The core of CLAMP is the unification of a learned "skill style" with environmental "logic." This is achieved by modeling the skill as a trajectory prior and the environment as a likelihood function.

The CLAMP framework architecture: Demonstrations are used to estimate prior parameters for a state-space trajectory prior, which is then combined with event likelihoods for MAP inference.

The trajectory prior is learned from human demonstrations using a Gaussian Process (GP) framework. Specifically, the authors utilize Linear Time-Varying Stochastic Differential Equations (LTV-SDEs) to represent the skill. This approach captures the heteroscedastic nature of human skills—where precision requirements vary throughout the task—while maintaining computational efficiency through sparsity. By modeling the skill as a linear stochastic dynamical system, the authors ensure the precision matrix is sparse, turning a heavy GP problem into a fast optimization.

3. Inference via Factor Graphs

To reproduce a skill, the robot must find the trajectory $\theta^*$ that maximizes the posterior probability $p(\theta | e) \propto p(e | \theta) p(\theta)$. Environmental constraints, such as obstacle avoidance and specific start/goal states, are encoded in the likelihood $p(e | \theta)$.

Factor graph representations of the inference process: (a) the basic GP prior, (b) the prior with start state constraints, and (c) the full model including environmental observations.

By leveraging the sparse structure of the LTV-SDE formulation, the MAP inference problem is mapped onto a factor graph. This allows for extremely fast optimization using standard solvers, enabling real-time planning and adaptation. Unlike previous two-step methods, CLAMP performs this unification in a single inference step, ensuring the final motion is both stylistically correct and safe.

4. Experimental Evaluation

The CLAMP framework was validated using a Kinova JACO 6-DOF robot arm on three manipulation tasks: box-opening, drawer-opening, and picking. The model was trained on kinesthetic demonstrations where a human guided the robot through the tasks.

Key results include:

  • Robust Generalization: The robot successfully performed tasks starting from initial states not encountered during training.
  • Dynamic Obstacle Avoidance: When unforeseen obstacles were introduced, CLAMP successfully deformed the trajectory to maintain safety while preserving the essential characteristics of the skill (e.g., the linear pull required for a drawer).
  • Computational Efficiency: The factor graph-based inference allowed for rapid convergence, supporting the potential for use in dynamic, real-world environments.

5. Conclusion

CLAMP offers a principled and efficient way to combine the benefits of LfD and MP. By treating skill reproduction as probabilistic inference, it provides a robust mechanism for robots to learn from humans and adapt to the complexities of the physical world. Future work will explore automated parameter tuning and more complex state representations to further enhance the system's versatility.

Summary updated

Abstract

In this paper, we present Combined Learning from demonstration And Motion Planning (CLAMP) as an efficient approach to skill learning and generalizable skill reproduction. CLAMP combines the streng...
BibTeX
@inproceedings{rana2017towards,
  author = {Muhammad Asif Rana and Mustafa Mukadam and Seyed Reza Ahmadzadeh and Sonia Chernova and Byron Boots},
  title = {{Towards Robust Skill Generalization: Unifying Learning from Demonstration and Motion Planning}},
  booktitle = {Conference on Robot Learning},
  year = {2017},
  url = {https://openreview.net/forum?id=PAgQiOfBLm4},
}
Figure