Research map / Reinforcement learning / Visual and representation RL
Shared Autonomy with IDA: Interventional Diffusion Assistance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Interventional Diffusion Assistance (IDA), a goal-agnostic value-based rule where a diffusion copilot intervenes only when its action beats the human pilot's action across all possible goals, evaluated in Reacher, Lunar Lander, and human-in-the-loop Lunar Lander.
Problem
Traditional shared autonomy keeps the copilot constantly active, which limits human autonomy and can hurt performance, and existing control-sharing methods require hand-tuned hyperparameters or goal knowledge.
Approach
A goal-agnostic intervention function uses an expert soft Q-function and marginalizes over possible goals; the copilot action is played only when it has higher expected value than the pilot action for all goals, with a diffusion copilot trained on goal-masked expert demonstrations.
Key results
- In Lunar Lander with 8 human participants, IDA achieved a 91.7% success rate versus 14.0% for pilot-only and 68.2% for copilot control, with significantly lower crash rate than pilot-only.
- In Reacher with noisy surrogate pilot, IDA improved hit rate from 1.6 to 3.3 targets per minute in the continuous goal space.
- Participants rated IDA as easiest and as preserving autonomy comparably to pilot-only control, and significantly better than copilot control (p<0.01).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IDA instead of pilot-only control | Reacher Continuous with noisy surrogate pilots · Reacher Continuous, noisy pilot | target hit rate (per minute): 1.6 -> 3.3 | Helped |
| use IDA instead of pilot-only control | Reacher Continuous with expert pilots · Reacher Continuous, expert pilot | target hit rate (per minute): 18.8 -> 18.8 | No clear effect |
| use IDA instead of pilot-only control | Lunar Lander noisy surrogate pilots · Lunar Lander, noisy pilot | success rate: 21% -> 83% | Helped |
| use IDA instead of Copilot (diffusion) | Lunar Lander laggy surrogate pilots · Lunar Lander, laggy pilot | success rate: 92% -> 100% | Helped |
| use IDA instead of human pilot-only control | human pilots in Lunar Lander · Human-in-the-Loop Lunar Lander | success rate: 14.0 (16.8) % -> 91.7 (4.9) % | Helped |
| use IDA instead of w/ Copilot | human pilots in Lunar Lander · Human-in-the-Loop Lunar Lander | success rate: 68.2 (9.1) % -> 91.7 (4.9) % | Helped |
| use IDA instead of w/ Copilot | human pilots in Lunar Lander · Human-in-the-Loop Lunar Lander | timeout rate: 13.8 (5.5) -> 0.1 (0.4) | Helped |
| use IDA instead of human pilot-only control | human pilots in Lunar Lander · Human-in-the-Loop Lunar Lander | crash rate: 85.8 (16.7) % -> 8.2 (5.1) % | Helped |
| use faux goals at inference (ablation) | IDA on Reacher Continuous with noisy pilots · Reacher Continuous, noisy pilot | target hit rate (per minute): 3.3 -> 2.9 | Hurt |
| train the expert Q function on a different goal distribution (ablation) | IDA on Reacher Continuous with noisy pilots · Reacher Continuous, noisy pilot | target hit rate (per minute): 3.3 -> 2.4 | Hurt |
| replace diffusion copilot with an MLP copilot (ablation) | interventional assistance on Lunar Lander, noisy pilots · Lunar Lander, noisy pilot | success rate: 83% -> 83% | No clear effect |
Surprising findings
- The diffusion copilot alone achieved 0 hit rate in Reacher, failing to acquire any targets even though it helped in Lunar Lander.
- Under copilot-only control, human participants experienced frequent timeouts (13.8%), showing the always-on copilot suppressed landings.
Limitations
- Requires an expert Q-function trained in a simulator, which the authors note is a limitation.
- Human-in-the-loop study uses only eight participants with no prior experience, one environment, and IDA still crashes 8.2% of the time.
Experimental setup
Datasets: expert demonstrations from SAC expert rollouts in Reacher expert demonstrations from SAC expert rollouts in Lunar Lander
Benchmarks: Reacher (continuous, linear, quadrant goal spaces) Lunar Lander (surrogate pilots) Lunar Lander human-in-the-loop
Metrics: hit rate per minute success rate crash rate timeout rate subjective ease, control, autonomy ratings Wilcoxon signed-rank p-values
Baselines: Pilot-only control Copilot (diffusion, gamma=0.2) Copilot (MLP) Intervention-Penalty (MLP) IA (MLP) IDA (FG) IDA (DS)
Builds on: DDPM soft actor-critic To the Noise and Back: Diffusion for Shared Autonomy On Optimizing Interventions in Shared Autonomy Teacher-Student Framework with value-based intervention
Closest papers by meaning
- Beyond Single Stationary Policies: Meta-Task Players as Naturally Superior Collaborators (NeurIPS 2024 · similarity 0.84)
- AI Agents Push Humans Out of the Loop (NeurIPS 2026 · similarity 0.83)
- Learning to Assist Humans without Inferring Rewards (NeurIPS 2024 · similarity 0.82)
- Designs for Enabling Collaboration in Human-Machine Teaming via Interactive and Explainable Systems (NeurIPS 2024 · similarity 0.82)
- AI Progress Should Be Measured by Capability-Per-Resource, Not Scale Alone: A Framework for Gradient-Guided Resource Allocation in LLMs (NeurIPS 2025 · similarity 0.82)
- Path Dependence under Adaptive AI Delegation (NeurIPS 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.