Research map / Reinforcement learning / Visual and representation RL
GoIRL: Graph-Oriented Inverse Reinforcement Learning for Multimodal Trajectory Prediction
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 GoIRL, a MaxEnt inverse reinforcement learning trajectory predictor that feeds vectorized lane-graph features into a grid reward via a feature adaptor, then decodes Bezier-parameterized trajectories from MCMC-sampled plans with refinement and probability fusion.
Problem
Multimodal trajectory prediction for surrounding agents in autonomous driving, where supervised behavior-cloning predictors suffer from modality collapse and poor generalization under drivable-area changes (covariate shift), and prior IRL predictors use rasterized BEV inputs that lose scene informat
Approach
A feature adaptor aggregates graph-based (lane and drivable-area) context features into a grid space so that a MaxEnt IRL reward and policy can be learned on vectorized context; sampled grid plans condition a hierarchical Bezier-curve trajectory decoder with a refinement stage, and MCMC plan frequencies are fused with classification probabilities.
Key results
- Single-model GoIRL reaches brier-minFDE_6 of 1.7957 on Argoverse test, versus 1.8584 for DSP and 1.8399 for HeteroGCN.
- Ensemble GoIRL reaches brier-minFDE_6 of 1.6947 on Argoverse test, the best among listed ensembles including Wayformer (1.7408).
- On nuScenes, GoIRL reports minADE_5 of 0.86 and minADE_10 of 0.75, outperforming P2T (1.45 and 1.16) and PGP (1.27 and 0.94).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the refinement module (ablation) | GoIRL · Argoverse test | minFDE_6: 1.1812 -> 1.1725 | Helped |
| add Bézier curve parameterization (ablation) | GoIRL · Argoverse test | minFDE_6: 1.1751 -> 1.1725 | Helped |
| add recurrent point-generation strategy (ablation) | GoIRL · Argoverse test | minFDE_6: 1.1764 -> 1.1725 | Helped |
| add MCMC-induced probability fusion (ablation) | GoIRL · Argoverse test | Brier score: 0.6310 -> 0.6232 | Helped |
| use GoIRL instead of HeteroGCN | Argoverse test single-model prediction · Argoverse test | brier-minFDE_6: 1.8399 -> 1.7957 | Helped |
| use GoIRL instead of Wayformer | Argoverse ensemble prediction · Argoverse test, ensemble | brier-minFDE_6: 1.7408 -> 1.6947 | Helped |
| use GoIRL instead of Goal-LBP | nuScenes prediction · nuScenes prediction benchmark | minADE_5: 1.02 -> 0.86 | Helped |
Limitations
- Explores only single-agent (target agent) forecasting; joint multi-agent prediction is future work.
- Failure cases include poor observations of static or slow agents, speed inaccuracies, and missed lane changes without explicit cues.
- IRL inner loop is computationally expensive, limiting scalability to larger state spaces.
Experimental setup
Datasets: Argoverse nuScenes
Benchmarks: Argoverse nuScenes
Metrics: brier-minFDE_6 minFDE_6 minADE_6 MR_6 Brier score minADE_5 minADE_10 minFDE_1 MR_5 MR_10
Baselines: LaneRCNN LaneGCN DSP HeteroGCN MultiPath++ SIMPL Wayformer P2T PGP MacFormer Goal-LBP
Builds on: LaneGCN DSP MaxEnt IRL Wayformer
Closest papers by meaning
- CarPlanner: Consistent Auto-regressive Trajectory Planning for Large-Scale Reinforcement Learning in Autonomous Driving (CVPR 2025 · similarity 0.88)
- Massively Scalable Inverse Reinforcement Learning in Google Maps (ICLR 2024 · similarity 0.87)
- Grounding Driving VLA via Inverse Kinematics (NeurIPS 2026 · similarity 0.85)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.85)
- Graph-Based Prediction and Planning Policy Network (GP3Net) for Scalable Self-Driving in Dynamic Environments Using Deep Reinforcement Learning (AAAI 2024 · similarity 0.85)
- UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning (NeurIPS 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.