Imitating Human Search Strategies for Assembly

Dennis Ehlers, Markku Suomalainen, Jens Lundell, Ville KyrkiICRA 2019 · DOI

TL;DR

This paper introduces a novel Learning from Demonstration (LfD) method that teaches robots to perform human-like search strategies for assembly tasks where precise alignment is difficult due to position uncertainty. By learning both the robot's movement dynamics and a probabilistic "exploration distribution" from human demonstrations, the system can generate effective search trajectories, significantly improving success rates in challenging 2D and 3D assembly scenarios.

What Problem Does This Paper Address?

Robotic assembly tasks, especially those involving small clearances like peg-in-hole or plug insertion, often fail when the robot's internal model of the environment doesn't perfectly match reality. Traditional methods rely on handcrafted exception strategies, which are tedious to create and don't generalize well. While LfD has shown promise in teaching robots specific skills, existing approaches typically assume a known or easily identifiable goal pose. This paper tackles the problem of teaching robots to search for a goal when its exact location or orientation is uncertain, mimicking how humans intuitively probe and explore an environment when they can't see precisely. This is crucial for robust assembly in the face of real-world uncertainties.

Methods

The core of the proposed method is a four-step framework:

  1. Learning the Exploration Distribution: The robot learns a probabilistic model (a Gaussian distribution) of where a human demonstrator searches. This distribution captures the area covered during the search, even if the goal's precise location is unknown.
  2. Learning Task Dynamics: Simultaneously, the system learns the robot's interaction dynamics (how forces and torques change with motion) from the demonstrations. This is crucial for compliant, force-aware searching.
  3. Creating the Search Trajectory: Two methods are proposed for generating a search trajectory that covers the learned exploration distribution:
  4. TSHIX (Trajectory Sampling from Human-Inspired Exploration): A novel sampling-based approach that samples points from the exploration distribution, finds an approximate shortest path through them (using a Traveling Salesman Problem solver), and then smooths the path.
  5. Ergodic Control: A deterministic method that aims to create a trajectory whose time-averaged behavior matches the desired exploration distribution.
  6. Executing the Search Trajectory: An impedance controller is used to execute the generated trajectory. Crucially, it incorporates the learned dynamics as superposed feed-forward forces, allowing for force-based searching.

Datasets

The paper does not introduce new datasets. Instead, it uses demonstrations collected from a KUKA LWR4+ robot performing two specific tasks:

  • 2D Peg-in-Hole Task: Demonstrations were collected for inserting a peg into a hole.
  • 3D Electrical Plug and Socket Task: Demonstrations were collected for inserting an electrical plug into a socket.

These tasks were chosen to evaluate the method in both simpler 2D and more complex 3D scenarios.

Key Insights and Ideas

  • Human-like probing is key: Robots can learn effective search strategies by imitating how humans explore uncertain environments, rather than relying solely on precise path following.
  • Learning dynamics enables force-based search: Incorporating learned interaction dynamics allows the robot to use forces to guide its search, which is essential for compliant contact tasks.
  • Sampling-based trajectory generation (TSHIX) offers flexibility: TSHIX allows for generating new trajectories on-the-fly if a previous one fails, unlike deterministic methods.
  • Data efficiency is achievable: The method demonstrates success with very few demonstrations (one for 2D, two for 3D), a significant advantage over many LfD approaches.
  • Demonstrations should capture variability: For the method to be effective, demonstrations need to cover the range of possible scenarios and dynamics encountered in the task, ideally from "worst-case" starting positions.

Key Results

  • Improved Success Rates: In the 2D peg-in-hole task, TSHIX achieved an 83% success rate compared to 67% for ergodic control. For the more complex 3D plug-in-socket task, TSHIX achieved a 67% success rate with two demonstrations, compared to only 20% with one demonstration and 0% for a random walk baseline.
  • Force-based search is beneficial: In the 2D peg-in-hole task, using superposed forces significantly improved success rates compared to executing trajectories without them.
  • TSHIX vs. Ergodic Control: TSHIX generally performed better, especially in scenarios where the ergodic trajectory might not fully cover the search area.
  • Sample size impact: For TSHIX, increasing the number of sampled points improved success rates up to a point (300 samples), after which it plateaued, suggesting diminishing returns and a trade-off with computational time.

Limitations and Future Work

  • "Learned what it has seen" limitation: Like most LfD methods, the robot can only learn strategies that were demonstrated. It cannot invent entirely new search behaviors.
  • Demonstration quality matters: The success of the method is highly dependent on the quality and coverage of the human demonstrations. Demonstrations need to capture the variability of the task.
  • Infeasibility in high dimensions: For extremely complex or high-dimensional tasks, even human teachers struggle, suggesting that some search problems might be inherently infeasible.
  • Future Directions:
    • Modeling high-dimensional searches as sequences of lower-dimensional ones.
  • Incorporating temporal correlations from human search into the exploration distribution.
  • Developing better teaching environments to ensure high-quality demonstrations.
  • Investigating how to improve trajectory following, potentially by better compensating for internal robot friction.

Summary updated

BibTeX
@inproceedings{ehlers2019imitating,
  author = {Dennis Ehlers and Markku Suomalainen and Jens Lundell and Ville Kyrki},
  title = {{Imitating Human Search Strategies for Assembly}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2019},
  doi = {10.1109/icra.2019.8793780},
}
Figure