Research map / Reinforcement learning / Visual and representation RL
Graph-Based Prediction and Planning Policy Network (GP3Net) for Scalable Self-Driving in Dynamic Environments Using Deep Reinforcement Learning
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 GP3Net, which combines a graph-based (Trajectron-style CVAE/GMM) trajectory predictor that builds future BEV occupancy masks with a PPO-trained state-encoder and policy for end-to-end driving in CARLA under traffic and weather shifts.
Problem
Autonomous driving motion planning in non-stationary environments, where imitation learning from expert demonstrations generalizes poorly to domain shifts and near-failure scenarios and existing RL planners do not model spatio-temporal interactions of traffic participants.
Approach
Feed PPO-trained driving policy with past BEV masks plus future occupancy masks generated from uncertainty-aware (Gaussian-patch) predictions of a spatio-temporal graph trajectory predictor (Trajectron-style), with the whole network updated end-to-end by RL.
Key results
- GP3Net achieves 82.5% success rate and 0.87 driving score in Town 03 vs. 80.2% and 0.85 for DiGNet (Full), and 92.5% SR in Town 01 and 93.1% in Town 02.
- In the weather generalization test on Towns 1 and 2, GP3Net reports success rates such as 96% (Wetnoon) and 88% (HardRainNoon) in Town 1, versus 84% and 80% for HIIL on the same conditions.
- On NoCrash, GP3Net's success rate drops from 94% on Town 2 train weather to 80% on new weather, with 96% vs. 90% end reach.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GP3Net instead of GAT | CARLA motion planning · Town 03 urban | success rate: 76.0 -> 82.5 | Helped |
| Use GP3Net instead of GAT | CARLA motion planning · Town 04 mixed | success rate: 71.0 -> 75.0 | Helped |
| Use GP3Net instead of GAT | CARLA motion planning · Town 05 urban | driving score: 0.87 -> 0.91 | Helped |
| Use GP3Net instead of GAT | CARLA motion planning · Town 06 highway | success rate: 78.2 -> 82.5 | Helped |
| Add the trajectory prediction module (ablation) | GP3Net · CARLA training | collision avoidance | Helped |
| Add the trajectory prediction module (ablation) | GP3Net · CARLA training after 2M steps | composed driving score | Helped |
| Use GP3Net instead of HIIL | CARLA weather robustness · Town 1 and Town 2, eight weather conditions | success rate: 96 vs 84; 95.9 vs 84; 97.3 vs 88; 92 vs 88; 92 vs 89; 94 vs 89; 96 vs 80; 88 vs 80; 93.3 vs 86; 94 vs 86; 98 vs 85; 94 v | Helped |
Surprising findings
- Without the trajectory prediction module, the PPO agent performed better up to about 2M training steps; only after that did the prediction-augmented model outperform it.
Limitations
- Gains over baselines are modest and the paper reports mixed results (e.g., Town 2 NoCrash success rate lower than Town 1; T1 SoftRainNoon ties HIIL)
- Trajectory predictor trained on data from a single town (Town 1) with a simple Gaussian-patch uncertainty model
- Prediction module only benefits training after about 2M steps, implying slow early learning
Experimental setup
Datasets: CARLA simulator roaming agent trajectory data (Town 1) CARLA Leaderboard scenarios CARLA NoCrash benchmark
Benchmarks: CARLA Leaderboard CARLA NoCrash
Metrics: success rate driving score end reach penalty factor collision rate infractions
Baselines: CILRS MLP GCN GAT DiGNet HIIL
Builds on: Trajectron Proximal Policy Optimization DiGNet HIIL
Closest papers by meaning
- Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.87)
- Driving in the Occupancy World: Vision-Centric 4D Occupancy Forecasting and Planning via World Models for Autonomous Driving (AAAI 2025 · similarity 0.86)
- IntersectionZoo: Eco-driving for Benchmarking Multi-Agent Contextual Reinforcement Learning (ICLR 2025 · similarity 0.86)
- NEST: A Neuromodulated Small-world Hypergraph Trajectory Prediction Model for Autonomous Driving (AAAI 2025 · similarity 0.86)
- Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules (AAAI 2025 · similarity 0.85)
- GoIRL: Graph-Oriented Inverse Reinforcement Learning for Multimodal Trajectory Prediction (ICML 2025 · similarity 0.85)
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.