Research map / Multi-agent systems and games / Human-AI and goal-conditioned
SPACeR: Self-Play Anchoring with Centralized Reference Models
Wei-Jer Chang, Akshay Rangesh, Kevin Joseph, Matthew Strong, Masayoshi Tomizuka, yihan hu, Wei Zhan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| anchor self-play to a pretrained tokenized reference model (SPACeR vs PPO) | self-play PPO · WOSAC val (vehicles) | composite realism: 0.710 -> 0.741 | Improved |
| anchor self-play to reference model | self-play PPO · WOSAC val (vehicles) | minADE: 12.725 -> 4.101 | Improved |
| add KL alignment to reference (Goal + KL vs Goal + LLH) (ablation) | SPACeR · WOSAC val (vehicles, anchoring ablation) | composite realism: 0.69 -> 0.73 | Improved |
| add log-likelihood reward only on top of task reward (ablation) | SPACeR · WOSAC val (vehicles, anchoring ablation) | composite realism: 0.70 -> 0.69 | No clear effect |
| add log-likelihood reward on top of KL + infraction penalties (ablation) | SPACeR · WOSAC val (vehicles, anchoring ablation) | composite realism: 0.74 -> 0.74 | No clear effect |
| remove explicit goal reward when anchored (KL + r_inf vs Goal + KL) (ablation) | SPACeR · WOSAC val (vehicles, anchoring ablation) | composite realism: 0.73 -> 0.74 | Improved |
| remove reference KL loss (ablation) | SPACeR · WOSAC val (pedestrians and cyclists) | composite realism: 0.607 -> 0.729 (adding KL) | Improved |
| add multi-action head (ablation) | SPACeR · WOSAC val (pedestrians and cyclists) | composite realism: 0.685 -> 0.729 | Improved |
| add goal-reaching weight (ablation) | SPACeR · WOSAC val (pedestrians and cyclists) | composite realism: 0.728 -> 0.729 | No clear effect |
| use weaker 0.3M reference model (ablation) | SPACeR · WOSAC val (vehicles) | composite realism: 0.636 reference -> 0.732 SPACeR | No clear effect |
| replace imitation model with SPACeR policy | CAT-K / SMART · WOSAC val (vehicles) | throughput (scenarios/s): 22.5 -> 211.8 | Improved |
| SPACeR vs CAT-K on composite realism | SPACeR · WOSAC val (vehicles) | composite realism: 0.766 -> 0.741 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SMART — SMART: Simultaneous Multi-Agent Recurrent Trajectory Prediction (ECCV 2020) · better on 3 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning to Drive in New Cities Without Human Demonstrations (NeurIPS 2026 · similarity 0.90)
- Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.90)
- Robust Autonomy Emerges from Self-Play (ICML 2025 · similarity 0.88)
- MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.88)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.87)
- Diffusion-Based Planning for Autonomous Driving with Flexible Guidance (ICLR 2025 · similarity 0.87)
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.
