TIGFlow-GRPO: Trajectory Forecasting via Interaction-Aware Flow Matching and Reward-Driven Optimization

Authors: Xuepeng Jing, Wenhuan Lu, Hao Meng, Zhizhi Yu, Jianguo Wei Date: 2026-03-26 Paper ID: openalex:2603.24936

Summary

TIGFlow-GRPO is a two-stage generative framework designed to improve human trajectory forecasting by integrating social and physical constraints into flow-based models. The first stage utilizes a Trajectory-Interaction-Graph (TIG) module to enhance the context encoding within a Conditional Flow Matching (CFM) predictor, capturing rich spatio-temporal interactions. The second stage introduces Flow-GRPO post-training, which converts the deterministic flow rollout into a stochastic SDE sampling process optimized by a composite reward that enforces view-aware social compliance and map-aware physical feasibility. Evaluations on ETH/UCY and SDD show that this approach achieves better accuracy and generates more behaviorally plausible, stable long-horizon trajectories.

Key Contributions

• Proposed TIGFlow-GRPO, a two-stage framework connecting Conditional Flow Matching trajectory generation with behavior-aware alignment.
• Introduced the Trajectory-Interaction-Graph (TIG) module within the CFM framework to effectively model fine-grained agent-agent and agent-scene interactions.
• Developed Flow-GRPO post-training to refine predictions by reformulating deterministic flow rollout as stochastic SDE sampling guided by composite rewards for social compliance and physical feasibility.
• Demonstrated improvements in accuracy, long-horizon stability, and behavioral realism on ETH/UCY and SDD datasets.

Limitations

The paper primarily focuses on connecting flow-based modeling with behavioral alignment, suggesting further exploration might be needed to fully analyze the impact of different reward components or the scalability to very large scenes.

Open Questions & Future Work

• statistical-distribution-vs-behavioral-alignment
• stochastic-rl-optimization-of-ode-flows
• long-horizon-error-accumulation-mitigation

Key Concepts

• trajectory-interaction-graph-tig: A graph-based module used within a Conditional Flow Matching framework to encode rich spatio-temporal interactions between agents and the scene for trajectory forecasting.
• flow-grpo-post-training: A post-training optimization technique that applies a Reward-Driven Optimization (GRPO) over the stochastic sampling process derived from a Continuous Normalizing Flow (CNF) to steer trajectory predictions towards desired behavioral outcomes.

Archivist Review

Two core reusable concepts (TIG and Flow-GRPO Post-training) were approved as they represent novel methodological contributions to combining flow-based generation with RL alignment. Two established open questions concerning the distribution-vs-behavior gap and RL optimization over flow dynamics were also approved, alongside a strong candidate on long-horizon error accumulation. The final proposed open question was rejected as it was too vague and listed as generic future work. The datasets ETH/UCY and SDD were approved as they are standard benchmarks in trajectory forecasting.

Approved Concepts

• Trajectory-Interaction-Graph (TIG): It explicitly models fine-grained agent-agent and agent-scene relations via a graph structure to enrich the conditional features for the flow-based model.
• Flow-GRPO Post-training: This novel two-stage approach explicitly aligns flow-based generative modeling with behavioral objectives (social compliance, physical feasibility) using Reinforcement Learning (GRPO) applied to the flow’s SDE formulation.

Approved Open Questions

• Bridging Distribution Fitting and Behavioral Alignment: Successfully bridging this gap is crucial for deploying trajectory forecasters in safety-critical applications like autonomous driving, where adherence to physical and social laws is paramount.
• Stochastic RL Optimization of ODE Flows: The ODE-to-SDE reformulation allows for the application of RL techniques to continuous generative models, but the stability and optimal control of this stochastic sampling for long-horizon, multimodal motion prediction warrant further dedicated research.
• Mitigating Long-Horizon Error Accumulation: Improving long-horizon stability is a primary bottleneck for autonomous system planning, as decision-making relies on accurate predictions far into the future.

Rejected Candidates

• [open_question] Richer Visual and VLM Integration (richer-visual-and-vlm-integration) - generic: This is presented as boilerplate future work rather than a specific, established open technical bottleneck.

Datasets

• ethucy
• sdd

Links

• Abstract
• PDF