arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

Agent‑Relative Goals Enable Real‑Time Interactive Driving Scenarios in Self‑Play | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
RoboticsEnglishPublished

Agent‑Relative Goals Enable Real‑Time Interactive Driving Scenarios in Self‑Play

November 20, 2026arXiv: 2611.24876v1

Researchers have shown that adding agent‑relative dynamic goals to a multi‑agent self‑play policy lets the system generate interactive driving behaviors such as cut‑offs, tail‑gating, and brake‑checking that were impossible with static map‑location goals. The updated framework conditions each agent’s policy on both a fixed world goal and a relative goal defined with respect to another agent’s position.

During training, 10 % of scenarios use agent‑relative goals while 90 % use static goals. This mix is achieved by sampling a categorical distribution over the four cardinal directions (front, back, left, right) and a distance distribution that favors longitudinal offsets. The policy is trained end‑to‑end without scripted trajectories or per‑scenario reward tuning.

When tested on the San Francisco map, the agent‑relative policy succeeded in 84.2 % of cut‑off roll‑outs, compared to only 14.6 % for a baseline that used only static goals. Similar performance gaps appear on other maps, confirming the benefit of the new goal representation.

Adding the relative goals barely affected core driving metrics. The collision percentage dropped by only 0.03 % and the off‑road tick percentage by 0.01 % relative to the baseline. No degradation was observed when evaluating the policy with static goals alone.

These results demonstrate that agent‑relative goals extend self‑play scenario generation without compromising baseline driving quality, offering a flexible tool for long‑tail autonomous‑vehicle testing.

Source paper

Learning to Interact in Multi-Agent Self-Play for Controllable Autonomous Driving Scenario Generation

RoboticsMachine Learning
View on arXiv