Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Maximal Adaptation, Minimal Guidance: Permissive Reactive Robot Task Planning with Humans in the Loop

Oz Gitelson, Satya Prakash Nayak, Ritam Raha, Anne-Kathrin Schmuck

TL;DR

We address the problem of persistent robot satisfaction of an infinite-horizon $LTL$ task $\varphi$ while a human pursues unknown latent objectives. The authors develop an online, autonomy-preserving framework built on maximal adaptation and a tunable feedback mechanism, realized through permissive strategy templates in $\omega$-regular games with parity-game reductions. The approach yields emergent cooperative behaviors and formal guarantees, demonstrated in a real gridworld manipulation task with a Panda and in Overcooked-AI, across diverse human-robot alignments and feedback settings. The main contributions are (1) a general HRlI framework not requiring fixed cooperation, (2) a formal template-based synthesis approach leveraging $\Pi_r$ and $\Pi_h$, and (3) an online adaptation/feedback loop that preserves human autonomy while ensuring persistent task progress.

Abstract

We present a novel framework for human-robot \emph{logical} interaction that enables robots to reliably satisfy (infinite horizon) temporal logic tasks while effectively collaborating with humans who pursue independent and unknown tasks. The framework combines two key capabilities: (i) \emph{maximal adaptation} enables the robot to adjust its strategy \emph{online} to exploit human behavior for cooperation whenever possible, and (ii) \emph{minimal tunable feedback} enables the robot to request cooperation by the human online only when necessary to guarantee progress. This balance minimizes human-robot interference, preserves human autonomy, and ensures persistent robot task satisfaction even under conflicting human goals. We validate the approach in a real-world block-manipulation task with a Franka Emika Panda robotic arm and in the Overcooked-AI benchmark, demonstrating that our method produces rich, \emph{emergent} cooperative behaviors beyond the reach of existing approaches, while maintaining strong formal guarantees.

Maximal Adaptation, Minimal Guidance: Permissive Reactive Robot Task Planning with Humans in the Loop

TL;DR

We address the problem of persistent robot satisfaction of an infinite-horizon task while a human pursues unknown latent objectives. The authors develop an online, autonomy-preserving framework built on maximal adaptation and a tunable feedback mechanism, realized through permissive strategy templates in -regular games with parity-game reductions. The approach yields emergent cooperative behaviors and formal guarantees, demonstrated in a real gridworld manipulation task with a Panda and in Overcooked-AI, across diverse human-robot alignments and feedback settings. The main contributions are (1) a general HRlI framework not requiring fixed cooperation, (2) a formal template-based synthesis approach leveraging and , and (3) an online adaptation/feedback loop that preserves human autonomy while ensuring persistent task progress.

Abstract

We present a novel framework for human-robot \emph{logical} interaction that enables robots to reliably satisfy (infinite horizon) temporal logic tasks while effectively collaborating with humans who pursue independent and unknown tasks. The framework combines two key capabilities: (i) \emph{maximal adaptation} enables the robot to adjust its strategy \emph{online} to exploit human behavior for cooperation whenever possible, and (ii) \emph{minimal tunable feedback} enables the robot to request cooperation by the human online only when necessary to guarantee progress. This balance minimizes human-robot interference, preserves human autonomy, and ensures persistent robot task satisfaction even under conflicting human goals. We validate the approach in a real-world block-manipulation task with a Franka Emika Panda robotic arm and in the Overcooked-AI benchmark, demonstrating that our method produces rich, \emph{emergent} cooperative behaviors beyond the reach of existing approaches, while maintaining strong formal guarantees.
Paper Structure (13 sections, 2 theorems, 1 equation, 3 figures, 1 table)

This paper contains 13 sections, 2 theorems, 1 equation, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Contribution
  4. Problem Setup
  5. Reactive Planning Domain
  6. Temporal Tasks as LTL formulas
  7. Problem Statement
  8. Adaptability and Feedback Mechanisms via Permissive Strategy Templates
  9. $\omega$-Regular Games
  10. Permissive Strategy Templates
  11. Adaptation and Feedback Mechanism
  12. Experiments
  13. Gridworld Block-Manipulation

Key Result

Proposition 1

Given a reactive planning domain $\mathcal{D}$ and an LTL formula $\varphi$ over $\mathit{AP}$, we can construct a parity game $\mathcal{G}=\left\langle \mathcal{D}',\Omega\right\rangle$ such that there is a bijective correspondence between the runs of $\mathcal{D}$ and the runs of $\mathcal{D}'$, a

Figures (3)

  • Figure 1: A simplified gridworld block-manipulation domain from our experimental setup, where a Franka Emika Panda robotic arm takes turns with a human to place blocks in a $3\times 3$ grid. The robot places blue blocks, while the human places red blocks. The top-left inset illustrates the robot's feedback to the human, suggesting the removal of the block in cell $(1,3)$ or $(2,2)$. A video of the experiment is available at https://youtu.be/61thSZDj5Ks.
  • Figure 2: An example of a partial reactive planning domain for turn-based human-robot interaction in a grid world. The robot controls the circle states, while the human controls the rectangle states. Each state contains a $3\times 3$ grid showing the current positions of the human-placed objects (red squares) and the robot-placed objects (blue circles). Directed edges represent possible actions leading to successor states. The robot's objective is to repeatedly reach states with majority-occupied cells where the placed objects are non-adjacent (i.e., no two occupied cells are neighbors), as illustrated in states $\mathtt{t1}$ and $\mathtt{t2}$. Green dashed edges denote live actions, which are the suggested actions by the strategy templates, while dotted edges denote sequences of actions that lead to the target states.
  • Figure 3: Satisfaction of human and robot recipe tasks over time in Overcooked-AI for (a) identical recipes, (b) incompatible recipes, and (c) compatible recipes. The plots show the proportion of runs satisfying each objective as well as frequency of feedback given.

Theorems & Definitions (9)

  • Example 1
  • Example 2
  • Example 3
  • Definition 1
  • Proposition 1
  • Proposition 2
  • Example 4
  • Example 5
  • Remark 1