Planning by Simulation: Motion Planning with Learning-based Parallel Scenario Prediction for Autonomous Driving

1. Problem Statement

Autonomous driving requires making safe, efficient, and human-like decisions in complex traffic environments. A critical challenge is the interdependence between prediction and planning:

• The Loop Gap: Traditional frameworks often treat prediction (forecasting other agents' trajectories) and planning (deciding the ego vehicle's path) as separate, sequential modules.
• The Flaw: Most prediction models are "agent-centric," meaning they predict other agents' futures based on current observations without considering how the ego vehicle's own planned actions will influence those agents.
• Consequence: This lack of mutual influence modeling leads to inaccurate predictions, which in turn causes the planner to make unsafe or suboptimal decisions (e.g., failing to anticipate that a nearby car will yield if the ego vehicle slows down).
• Limitations of Existing Search Methods: While Monte Carlo Tree Search (MCTS) is used for planning, existing implementations often suffer from:
  1. Redundant Computation: They re-encode inputs every time the observation window slides, failing to reuse past computations.
  2. Discrete Action Sets: They rely on finite, discrete actions (e.g., specific yaw rates), creating a gap between the planned trajectory and the physically continuous reality.
  3. Unrealistic Branches: Tree expansion can generate branches that violate physical laws.

2. Methodology: Planning by Simulation (PS)

The authors propose PS, a novel framework that integrates learning-based parallel scenario prediction with MCTS to create a cyclic, interactive planning loop.

Core Architecture

The framework operates as an iterative simulation where the ego vehicle's potential actions and other agents' reactions are inferred simultaneously.

• MCTS Foundation: The planner uses MCTS to explore the space of possible futures. The tree search consists of four steps:

  1. Selection: Choosing nodes to explore based on the Upper Confidence Bound (UCB) to balance exploration and exploitation.
  2. Expansion: Adding new nodes by generating new ego actions and predicting corresponding agent reactions.
  3. Simulation (Rollout): Simulating future scenes to a fixed depth to evaluate the cost of a trajectory.
  4. Backpropagation: Updating node statistics (values and visit counts) based on simulation results.

• Ego Vehicle Transition (Frenet Frame):

  • Instead of using a single kinematic model, the ego vehicle's state is updated using the Frenet frame relative to the road centerline.
  • Target speeds are sampled, and lateral displacements are generated via polynomial fitting (quintic/quartic).
  • A cost function (based on Bellman's principle) selects the optimal path, ensuring smoothness and kinematic feasibility before converting back to Cartesian coordinates for the tree node.

• Other Agents Transition (Query-Centric Prediction):

  • Key Innovation: The system uses a Query-Centric Trajectory Prediction model. Unlike agent-centric models that require re-encoding the entire scene when the ego moves, this model encodes the scene relative to a "query" (the new state).
  • Efficiency: By treating the new ego state as a query and reusing the encoded map and historical agent embeddings (Keys/Values), the model can efficiently predict how all other agents will react to the ego's specific action without redundant computation.
  • Parallelism: This allows the system to simulate multiple future scenarios in parallel, where the ego's plan and the agents' reactions are co-inferred.

• Cost Function:
  The evaluation metric is a weighted linear combination of:

  1. Efficiency: Deviation from the speed limit.
  2. Comfort: Smoothness of acceleration and steering (jerk minimization).
  3. Safety: Penalization of hard braking/acceleration and collision risk (modeled via sigmoid functions on distance to other agents).

3. Key Contributions

1. Interactive Planning Framework: A data-driven framework that explicitly models the mutual influence between the ego vehicle's planning and other agents' predictions, resulting in more robust decision-making than pure end-to-end or rule-based approaches.
2. Iterative Cyclic Structure: A novel architecture where trajectory prediction guides the planning search, and the planning search (ego actions) simultaneously updates the prediction inputs, creating a closed-loop reasoning process.
3. Efficient Query-Centric Simulation: A reusable prediction network based on coordinate transformation (query-centric) that enables efficient, parallel simulation of tree-like future scenarios, avoiding the computational redundancy of sliding-window re-encoding.
4. Frenet-MCTS Integration: Combining Frenet frame path generation with MCTS to ensure kinematic feasibility and smooth action generation, overcoming the limitations of discrete action sets.

4. Experimental Results

The method was evaluated on the Argoverse 2 dataset across four challenging urban scenarios:

1. Opposing Left Turns: Handling uncertain obstacles while maintaining speed limits.
2. Lane Changing: Merging into gaps between fast-approaching rear and sluggish front vehicles.
3. Intersection Merging: Navigating when a target vehicle merges into the ego's lane.
4. Overtaking: Reacting to sudden stops and complex interactions on crowded roads.

Performance Metrics:

• Completion Time (C.T.): PS achieved the fastest average completion time (3.91s) compared to baselines like PS-MAE (5.04s) and PS-FJMP (4.19s).
• Average Velocity (A.V.): PS maintained a high average velocity (16.49 m/s) while adhering to constraints, outperforming methods that either slowed down too much or violated speed limits.
• Collision Distance (C.D.): PS maintained a safe distance (3.41m), demonstrating superior safety compared to baselines that resulted in collisions or near-misses in ablation studies.
• Ablation Studies:
  • PS-Rule (Rule-based prediction): Failed to adapt to dynamic interactions, leading to collisions or inefficient paths.
  • PS-Niter (Single inference): Showed that running prediction only once per action (without cyclic updates) degrades performance, proving the necessity of the iterative loop.
  • PS-Fixed (Fixed action set): Resulted in higher speeds but unsafe maneuvers, highlighting the need for adaptive action generation.

5. Significance

This paper addresses a fundamental bottleneck in autonomous driving: the separation of prediction and planning. By treating planning as a simulation problem where the ego vehicle's actions are part of the prediction input, the PS framework achieves:

• Higher Robustness: It anticipates how other drivers will react to the ego vehicle, leading to safer interactions.
• Computational Efficiency: The query-centric approach allows for deep search trees without the prohibitive computational cost of re-encoding scenes at every step.
• Human-Like Behavior: The system generates smooth, efficient, and socially compliant trajectories that mimic human negotiation in traffic.

The work suggests a future direction for autonomous driving where large-scale search (like MCTS) is tightly coupled with deep learning predictors to enable "reasoning" before acting, similar to the "Chain of Thought" in large language models.