# SLOT-V: Supervised Learning of Observer Models for Legible Robot Motion Planning in Manipulation

**Authors:** Sebastian Wallkotter, Mohamed Chetouani, Ginevra Castellano  
**arXiv:** [2210.1412](https://arxiv.org/abs/2210.1412)  
**Date:** 2022-10-04  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2210.1412v1/

## Key points

- The paper introduces SLOT-V, a supervised learning framework for training observer models in legible robot motion planning.
- Legibility refers to how easily a human observer can understand the goal of a robot based on its motion trajectory.
- Existing planners often rely on hand-crafted or demonstration-based observer models.
- SLOT-V proposes learning these observer models in a supervised manner using the same data format commonly used for evaluating existing approaches.
- SLOT-V can accurately predict various hand-crafted observer models, indicating its hypothesis space encompasses existing models.
- SLOT-V's performance remains good in environments with unseen goal configurations and/or goal counts, highlighting its generalization ability.
- Compared to an existing Inverse Reinforcement Learning (IRL) approach, SLOT-V learns better observer models with less data, suggesting superior sample efficiency.
- SLOT-V is capable of learning viable observer models that lead to more legible trajectories.
- Improved legibility has the potential to enhance human-robot interaction by facilitating better understanding and transparency between humans and robots.

## Summary

The paper introduces SLOT-V, a supervised learning framework for training observer models in the context of legible robot motion planning. Legibility refers to how easily a human observer can understand the goal of a robot based on its motion trajectory. Existing planners often rely on hand-crafted or demonstration-based observer models to estimate the quality of trajectory candidates. SLOT-V proposes learning these observer models in a supervised manner using the same data format commonly used for evaluating existing approaches. The authors demonstrate the generality of SLOT-V by applying it to a Franka Emika robot in a simulated manipulation environment. They show that SLOT-V can accurately predict various hand-crafted observer models, indicating that its hypothesis space encompasses existing models. Furthermore, the authors showcase SLOT-V's ability to generalize by testing its performance in environments with unseen goal configurations and/or goal counts. The trained model continues to perform well, highlighting its robustness and adaptability. To evaluate SLOT-V's sample efficiency and performance, the authors compare it against an existing Inverse Reinforcement Learning (IRL) approach. The results indicate that SLOT-V learns better observer models with less data, suggesting its superior sample efficiency. Overall, the findings suggest that SLOT-V is capable of learning viable observer models which leads to more legible trajectories; this improved legibility has potential to enhance human-robot interaction by facilitating better understanding and transparency between humans and robots.