Dreaming to Assist: Learning to Align with Human Objectives for Shared Control in High-Speed Racing
DOI: 10.48550/arxiv.2410.10062
archive: archived pipeline: cataloged verified
Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)
Summary
This paper addresses the challenge of effective human-robot team coordination in high-speed, dynamic environments like competitive racing, where explicit communication is impractical. The core problem is enabling a robotic assistive agent to infer a human driver’s tactical objectives (e.g., overtaking vs. staying behind) and provide control assistance that aligns with those intents. Existing shared-control methods often split authority along predefined boundaries or fail to account for the multimodal nature of human decision-making. The authors propose **DREAM2ASSIST**, a framework that uses a rich world model to infer human intent and an assistive agent to provide expert-aligned assistance, aiming to improve safety and performance while respecting the driver’s goals. The method employs a Model-Based Reinforcement Learning (MBRL) paradigm centered on a Recurrent State-Space Model (RSSM). This model jointly learns the physical dynamics of the environment and the human’s latent intent. To train the system, the authors use a "fictitious co-play" approach, generating a population of synthetic human drivers with distinct, mutually exclusive objectives (such as "pass" vs. "stay" or "left" vs. "right"). The RSSM is trained to predict observations, rewards, and the human’s intent. The assistive agent’s reward function is shaped to encourage alignment with the inferred intent, combining task-specific rewards (collision avoidance, track progress), an intervention penalty to minimize unnecessary control overrides, and an expert alignment term that rewards actions consistent with the optimal policy for the inferred human objective. Experiments were conducted in the CARLA simulator using two racing scenarios: a straightaway and a complex hairpin turn. The DREAM2ASSIST agent was evaluated against baselines including a standard Dreamer model without intent inference and an adversarial imitation learning baseline. Results demonstrated that DREAM2ASSIST significantly outperformed baselines in both track progress and collision avoidance. Specifically, in the challenging hairpin scenario, the intent-conditioned agent successfully disentangled distinct human modes, providing appropriate lateral or longitudinal corrections to support overtaking or staying behind, whereas baselines often failed to distinguish between modes, leading to aggressive off-track driving or ineffective assistance. The intent-conditioning mechanism allowed the robot to adhere to human preferences, resulting in improved team performance compared to synthetic humans acting alone. The significance of this work lies in demonstrating that explicit intent inference within a world model enables fluid, adaptive shared control in high-stakes domains. By conditioning assistance on inferred human objectives, the system avoids conflicts between human and robot actions, which are common in traditional shared-control schemes. The findings suggest that theory-of-mind-inspired models can enhance human-robot collaboration by allowing machines to reason about human goals and provide context-aware support. The authors note limitations, including the use of synthetic humans rather than real drivers, and suggest future work on generalizing to real-world human-robot teams and estimating optimal rewards without privileged model access.
Key finding
An assistive agent that infers human intent using a recurrent state-space model outperforms baseline methods by improving task performance and reducing collisions while respecting the driver's specific tactical objectives.
Methodology
simulation_modeling
Provenance
The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | author_sweep | — | — | 2 | 2026-05-28 |
| archive | success | canonical_url | — | — | 1 | 2026-06-04 |
| extract | success | cached | — | — | 3 | 2026-06-10 |
| clean | success | clean | — | — | 1 | 2026-06-04 |
| chunk | success | chunk | — | — | 1 | 2026-06-04 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-04 |
| enrich | success | — | — | — | 1 | 2026-05-28 |
| promote | success | — | — | — | 1 | 2026-06-04 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 2 | 2026-06-10 |
| tag | success | vector_similarity | — | — | 15 | 2026-06-11 |
| verify | success | — | — | — | 2 | 2026-06-10 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.
Information type
What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).
- Theoretical Contribution: computational model