Multi-Turn Trajectory Refinement in Autonomous Driving

• The paper introduces multi-turn trajectory refinement as a method that iteratively improves initial path proposals through sequential updates leveraging real-time feedback and interaction modeling.
• It employs techniques like neural feature refinement, diffusion-based denoising, and reinforcement learning to optimize trajectories for enhanced safety and scene compliance.
• Empirical results on benchmarks show significant reductions in displacement errors and improvement in safety metrics by consistently refining trajectories in dynamic, real-world scenarios.

Multi-turn trajectory refinement in autonomous driving refers to the class of planning and prediction algorithms that iteratively improve initial trajectory proposals through multiple refinement stages, leveraging new information, interaction modeling, or environmental feedback at each turn. This paradigm is motivated by the need for high-precision, scene-compliant, and robust path generation in complex, dynamic traffic scenarios, where single-pass (one-shot) methods often fail to capture interaction nuances, topological constraints, or to escape local minima. Multi-turn frameworks are broadly instantiated in both prediction and planning systems, including multi-agent forecasting, closed-loop control optimization, and end-to-end learning architectures.

1. Core Principles and Motivation

Multi-turn trajectory refinement extends conventional trajectory prediction and planning by introducing sequential correction or enhancement mechanisms. Each iteration—also called a refinement turn or stage—adjusts the agent’s path proposal according to evolving context, refined interaction features, or direct feedback from downstream evaluators. This approach can model spatio-temporal dependencies, agent interactions, and kinodynamic feasibility more effectively than single-pass designs. Empirical results across diverse benchmarks confirm that iterative refinement yields systematic improvements in core safety, comfort, and compliance metrics (Xiao et al., 6 Jul 2025, Yin et al., 21 Nov 2025, Li et al., 30 Jan 2026, Jiao et al., 5 Aug 2025, Jie et al., 2022, Sun et al., 11 Sep 2025).

This multi-turn principle encompasses several computational realizations:

• Iterative neural feature refinement or update (e.g., stacked attention layers);
• Gradient-based optimization in factor graphs or variational inference;
• Closed-loop reinforcement learning with cross-turn feedback or rewards;
• Diffusion-based coarse-to-fine denoising;
• Recursive multi-granularity temporal decoding.

2. Algorithmic Methodologies

2.1 Multi-Agent Trajectory Refinement with Topological Attention

Approaches such as SRefiner introduce explicit modeling of agent-agent and agent-lane topological relationships via soft-braid attention. This method processes the initial prediction from a base predictor and, across multiple refinement iterations (typically three), alternates between:

• Extracting updated spatio-temporal topological cues (soft-braid topology) from current predictions;
• Applying trajectory–trajectory and trajectory–lane soft-braid attention, infusing these topological features into agent embeddings;
• Predicting and composing incremental corrections to agent trajectories.

Each turn leverages increasingly precise context, improving both global consistency and lane adherence (Xiao et al., 6 Jul 2025).

2.2 Diffusion-Based Coarse-to-Fine Refinement

DiffRefiner demonstrates a two-stage framework combining a transformer-based proposal decoder (generating K coarse trajectory proposals with anchor offsets) and a conditional diffusion refiner. The diffusion stage iteratively denoises trajectory candidates in the presence of scene context, via global and local semantic interaction modules. Even one or two refinement steps often suffice due to high-quality proposal anchors, resulting in significant performance gains over single-pass discriminative models (Yin et al., 21 Nov 2025). Conditioning on both anchor and local scene features ensures compliance with map regions and real-time adaptability.

2.3 RL-Based Multi-Turn Interactive Refinement

MTDrive integrates multi-modal LLMs (MLLMs) with reinforcement learning in a closed-loop, multi-turn paradigm. At each turn, the model proposes a candidate trajectory, receives structured feedback from a perception-decision-metrics agent regarding collisions, drivable area violations, and time-to-collision, and incorporates this feedback into subsequent trajectory generations. The Multi-Turn Group Relative Policy Optimization (mtGRPO) algorithm normalizes rewards across turns for efficient credit assignment, leading to iterative improvement and strong empirical gains on NAVSIM benchmarks (Li et al., 30 Jan 2026).

2.4 Adversarial and Scalarization-Free Multi-Objective Refinement

EvaDrive coordinates the evolution of both trajectory generator and critic in a vector-valued, multi-round adversarial game. The generator alternates between autoregressive intent modeling and diffusion-based spatial refinement, while the critic outputs untied vector-valued rewards for multiple planning objectives (e.g., safety, efficiency, comfort). A Pareto frontier selection mechanism identifies non-dominated trajectories at each round, guiding exploration and diversity in subsequent refinements and ensuring the avoidance of scalarization bias (Jiao et al., 5 Aug 2025).

2.5 Recursive Multi-Granularity and Incremental Kinematic Refinement

MGTraj employs recursive refinement networks (RRNs) at successive temporal granularities, starting with a coarse, goal-driven interpolation and refining towards finer temporal scales. Each granularity-specific RRN predicts corrections in state and velocity, with shared parameterization across stages. This decomposes global planning uncertainty and facilitates synthesis of both tactical and fine-grained behavior (Sun et al., 11 Sep 2025).

Traditional optimization-based planning, as in Gaussian process and incremental refinement systems, decouples path and speed planning for kinodynamic feasibility. Path is generated using a GP prior in Frenét space with boundary and obstacle constraints, followed by s-t graph-based speed profile search; the complete trajectory is incrementally refined to enforce lateral acceleration and curvature constraints in dynamic scenarios. This loop continues until all kinodynamic constraints are satisfied, typically within a few refinement steps (Jie et al., 2022).

3. Spatio-Temporal Interaction and Topological Modeling

A distinguishing hallmark of advanced multi-turn refiners lies in their encoding of dynamic spatial interactions and topological constraints:

• Soft-braid topology (SRefiner) encodes agent-agent and agent-lane relative geometry, time-aligned intersection points, relative velocities, and orientations, applied via multi-head cross-attention (Xiao et al., 6 Jul 2025).
• Semantic interaction modules (DiffRefiner) integrate sentence-level and pixel-level environmental context—such as drivable area, crosswalks, and dynamic agents—into each iterative denoising stage for fine scene compliance (Yin et al., 21 Nov 2025).
• In adversarial RL (EvaDrive), Pareto-guided refinement ensures agents optimize across a vector of objectives without collapsing heterogeneous criteria (e.g., safety vs. comfort) into a single scalar reward (Jiao et al., 5 Aug 2025).
• Real-time factor-graph methods update path factors only at constraint-violating points, ensuring minimal computational overhead and accurate enforcement of critical kinodynamic limits (Jie et al., 2022).

4. Training, Supervision, and Optimization

Across frameworks, supervision and optimization employ staged or multi-iteration losses:

• Mode-wise winner-takes-all (WTA) selection and huber or L1 regression losses at each iteration promote the best mode and progressively reduce displacement error, as in SRefiner and DiffRefiner (Xiao et al., 6 Jul 2025, Yin et al., 21 Nov 2025).
• MTDrive’s mtGRPO normalizes per-turn RL rewards, jointly optimizing with a formatting-integrity component to align outputs with expected structural templates (Li et al., 30 Jan 2026).
• Recursive RRNs in goal-guided models use shared-parameter transformers with joint L2 losses on position and velocity, regularized via auxiliary prediction heads (Sun et al., 11 Sep 2025).
• Factor-graph approaches update only locally violated constraints per iteration, maintaining computational tractability (Jie et al., 2022).

5. Empirical Evaluation and Benchmarking

Quantitative results across major datasets demonstrate that multi-turn refinement delivers consistent and significant improvements relative to single-pass or baseline methods.

Method	Benchmark	Metric	Single-Pass	Multi-Turn (Best)	Relative Gain
SRefiner	Argoverse v2	avgMinFDE (m)	1.642	1.477	–10.1%
SRefiner	INTERACTIONS	minJointFDE (m)	0.630	0.579	–8.1%
DiffRefiner	NAVSIM v2	EPDMS	86.0*	87.4	+1.4
MTDrive	NAVSIM	PDMS	83.7	91.1†	+8.8
EvaDrive	NAVSIM v1	PDMS	88.1**	94.9	+7.7

*Best previous baseline, **DiffusionDrive, †mtGRPO (kinematic setting)

Notable findings include:

• Multi-turn soft-braid attention in SRefiner yields 5–12% relative reduction in displacement errors and miss rates (Xiao et al., 6 Jul 2025).
• Iterative RL methods surpass both single-turn and diffusion models, even exceeding human performance in closed-loop safety metrics (Li et al., 30 Jan 2026).
• Coarse-to-fine and Pareto-guided multi-stage refinement ensures solution diversity and robustness to objective trade-offs (Jiao et al., 5 Aug 2025, Yin et al., 21 Nov 2025).

6. Practical Implementation Details and System Considerations

Multi-turn refiners are deployed in both real-time and simulation settings:

• GP- and factor-graph–based incremental planners are implemented in C++ with efficient data structures (GTSAM, Bayes tree), enabling 20–58 Hz operation on commodity CPU hardware (Jie et al., 2022).
• CNN/Transformer-based multi-turn agents are trained using large-scale, closed-loop simulated datasets (NAVSIM, Bench2Drive, Argoverse), with batch sizes up to 512 and multi-stage fine-tuning for both proposal and refinement modules (Yin et al., 21 Nov 2025, Li et al., 30 Jan 2026).
• System-level optimizations—such as inter-process streaming serialization and in-GPU tensor caches—are employed within RL pipelines to achieve up to 2.5× throughput improvements during training (Li et al., 30 Jan 2026).
• Each framework typically limits the number of refinement turns to three to six, as empirical analysis shows diminishing gains beyond this range (Xiao et al., 6 Jul 2025, Yin et al., 21 Nov 2025).

7. Limitations, Variants, and Future Extensions

Several limitations remain:

• Some methods, such as MTDrive, rely on privileged PDM agents or ground-truth perception during feedback, which may not be available in real-world deployment (Li et al., 30 Jan 2026).
• Modeling of feedback or reward assignment is critical: sparse or delayed rewards can impede credit assignment unless addressed by mechanisms such as mtGRPO.
• The sim-to-real gap persists, as interactive feedback in simulation may not precisely translate to physical perception and actuation constraints.

Potential directions for future refinement include:

• Integration of fully end-to-end perception so that refinement modules directly reason about raw sensor input rather than map-based or handcrafted feature constraints (Li et al., 30 Jan 2026).
• Expansion to multi-agent scenarios and generalization to additional sequential-decision tasks beyond trajectory planning, such as robotic manipulation (Sun et al., 11 Sep 2025).
• Replacement or augmentation of hand-crafted evaluators with learned or human-in-the-loop reward models, and extension to other simulators such as CARLA and Waymax.

Multi-turn trajectory refinement has established itself as a central principle for top-performing autonomous driving systems, consistently driving progress in trajectory precision, safety compliance, and generalization under challenging interactive and dynamic settings (Xiao et al., 6 Jul 2025, Yin et al., 21 Nov 2025, Li et al., 30 Jan 2026, Jiao et al., 5 Aug 2025, Jie et al., 2022, Sun et al., 11 Sep 2025).