Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Shared Autonomy with IDA: Interventional Diffusion Assistance

Brandon J. McMahan, Zhenghao Peng, Bolei Zhou, Jonathan C. Kao

TL;DR

The paper tackles preserving human autonomy in shared autonomy by introducing Interventional Diffusion Assistance (IDA), a diffusion-based copilot trained on expert demonstrations with goal masking. It uses a trajectory-based, goal-agnostic intervention function to intervene only when the copilot outperforms the human across all goals, and provides theoretical guarantees on performance. Empirical evaluation in Reacher and Lunar Lander, including human-in-the-loop experiments, shows IDA can exceed pilot-only and copilot baselines while maintaining or enhancing user autonomy. The approach is modular and hyperparameter-free, with practical implications for safer, more capable human-AI collaborative control in high-dimensional tasks.

Abstract

The rapid development of artificial intelligence (AI) has unearthed the potential to assist humans in controlling advanced technologies. Shared autonomy (SA) facilitates control by combining inputs from a human pilot and an AI copilot. In prior SA studies, the copilot is constantly active in determining the action played at each time step. This limits human autonomy and may have deleterious effects on performance. In general, the amount of helpful copilot assistance can vary greatly depending on the task dynamics. We therefore hypothesize that human autonomy and SA performance improve through dynamic and selective copilot intervention. To address this, we develop a goal-agnostic intervention assistance (IA) that dynamically shares control by having the copilot intervene only when the expected value of the copilot's action exceeds that of the human's action across all possible goals. We implement IA with a diffusion copilot (termed IDA) trained on expert demonstrations with goal masking. We prove a lower bound on the performance of IA that depends on pilot and copilot performance. Experiments with simulated human pilots show that IDA achieves higher performance than pilot-only and traditional SA control in variants of the Reacher environment and Lunar Lander. We then demonstrate that IDA achieves better control in Lunar Lander with human-in-the-loop experiments. Human participants report greater autonomy with IDA and prefer IDA over pilot-only and traditional SA control. We attribute the success of IDA to preserving human autonomy while simultaneously offering assistance to prevent the human pilot from entering universally bad states.

Shared Autonomy with IDA: Interventional Diffusion Assistance

TL;DR

The paper tackles preserving human autonomy in shared autonomy by introducing Interventional Diffusion Assistance (IDA), a diffusion-based copilot trained on expert demonstrations with goal masking. It uses a trajectory-based, goal-agnostic intervention function to intervene only when the copilot outperforms the human across all goals, and provides theoretical guarantees on performance. Empirical evaluation in Reacher and Lunar Lander, including human-in-the-loop experiments, shows IDA can exceed pilot-only and copilot baselines while maintaining or enhancing user autonomy. The approach is modular and hyperparameter-free, with practical implications for safer, more capable human-AI collaborative control in high-dimensional tasks.

Abstract

The rapid development of artificial intelligence (AI) has unearthed the potential to assist humans in controlling advanced technologies. Shared autonomy (SA) facilitates control by combining inputs from a human pilot and an AI copilot. In prior SA studies, the copilot is constantly active in determining the action played at each time step. This limits human autonomy and may have deleterious effects on performance. In general, the amount of helpful copilot assistance can vary greatly depending on the task dynamics. We therefore hypothesize that human autonomy and SA performance improve through dynamic and selective copilot intervention. To address this, we develop a goal-agnostic intervention assistance (IA) that dynamically shares control by having the copilot intervene only when the expected value of the copilot's action exceeds that of the human's action across all possible goals. We implement IA with a diffusion copilot (termed IDA) trained on expert demonstrations with goal masking. We prove a lower bound on the performance of IA that depends on pilot and copilot performance. Experiments with simulated human pilots show that IDA achieves higher performance than pilot-only and traditional SA control in variants of the Reacher environment and Lunar Lander. We then demonstrate that IDA achieves better control in Lunar Lander with human-in-the-loop experiments. Human participants report greater autonomy with IDA and prefer IDA over pilot-only and traditional SA control. We attribute the success of IDA to preserving human autonomy while simultaneously offering assistance to prevent the human pilot from entering universally bad states.
Paper Structure (22 sections, 9 theorems, 58 equations, 6 figures, 4 tables, 1 algorithm)

This paper contains 22 sections, 9 theorems, 58 equations, 6 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Notation and Problem Formulation.
  5. Training an Expert Policy
  6. Training a Diffusion Copilot
  7. Trajectory-based Goal-agnostic Value Intervention
  8. Theoretical Guarantees on the Performance of IA
  9. Results
  10. Experimental Setup
  11. Reacher Experiments
  12. Lunar Lander
  13. Lunar Lander with Human-in-the-loop Control
  14. Conclusion and Discussion
  15. Acknowledgments
  16. ...and 7 more sections

Key Result

Theorem 1

Let $J(\pi) = \mathbf{E}_{s_0 \sim d_0, a_t \sim \pi(\cdot \mid s_t), s_{t+1} \sim P(\cdot \mid s_t, a_t)}[\sum_{t=0}^\infty \gamma^t r(s_t,a_t)]$ be the expected discounted return of following a policy $\pi$. Then, the performance following the Interventional Assistance policy (or behavior policy)

Figures (6)

  • Figure 1: Overview of Interventional Assist Framework for control sharing. (a) Prior works perform shared autonomy by passing human actions to a copilot javdani2015sharedyoneda2023noisereddy2018sharedjeon2020shared. The copilot then plays an action, e.g., by selecting a feasible action closest to the user suggestion reddy2018shared or through diffusion yoneda2023noise. (b) In this work, we design an intervention function that plays either the human pilot's action, $a_p$, or the copilot's action, $a_c$, based on their goal-agnostic advantages.
  • Figure 2: Reacher experiments. (a) Continuous Reacher environment. (b) Laggy pilot experiments as the number of possible goals varies. IDA performance slightly decreases as the number of possible goals increase, but it never significantly underperforms the pilot. The copilot is significantly worse than both the laggy pilot and IDA. (c) Noisy pilot experiments. IDA outperforms the pilot and copilot.
  • Figure 3: Analysis of copilot advantages during intervened states. (a) Characterization of intervention during the noisy pilot control. The left plot shows the copilot advantage, which is generally higher for corrupted (random) actions compared to pilot actions. When quantifying the number of intervened states, we see IDA intervenes more when the corrupted actions are taken. (b) Same as (a) but for the laggy pilot. (c) Example intervened states. In the top panel, the copilot prevents flipping. In the bottom panel, the copilot action helps to make a graceful landing.
  • Figure 4: Participants rated IDA as the easiest. Participants subjectively rated IDA as achieving a similar level of autonomy to pilot only control but significantly better than copilot control.
  • Figure A.1: Intervention tends to occur near the start of trajectories to stabalize the rocket and then again near the end of trajectories to assist the touch down.
  • ...and 1 more figures

Theorems & Definitions (16)

  • Theorem 1
  • Lemma 1
  • proof
  • Lemma 2: Policy difference lemma
  • Lemma 3
  • proof
  • Lemma 4
  • proof
  • Lemma 5: State Distribution Difference Bound
  • proof
  • ...and 6 more