Deciding When to Trust the Expert: VLM-Augmented Policy Switching for Safe Driving (VLAPS)

Qizhao Chen, Debajyoti Chakrabarti

CS 269 Course Research Project · Fall 2025
Vision-Language Models Imitation Learning TD Learning Motion Planning Autonomous Driving

Overview

Online reinforcement learning can require unsafe exploration, while imitation learning depends on large expert datasets and may struggle under distribution shift. VLAPS replaces continuous human supervision with a vision-language model that compares the short-horizon behavior of a learned policy and an expert controller and decides which one should control the vehicle.

The expert combines RRT* local planning with PD tracking. The VLM decision is used not only for policy switching during training, but also as supervision for behavior cloning and TD-style value learning.

VLAPS pipeline: expert and learner policies generate candidate trajectories, which are visualized and compared by a VLM that determines execution and learning supervision.
VLAPS pipeline. Expert and learner policies generate candidate trajectories, which are visualized and compared by a VLM that determines execution and learning supervision.

RRT* + PD Expert

The expert policy builds a LiDAR-derived ego-centric occupancy grid, plans a local collision-free path using RRT*, and tracks a lookahead waypoint using PD steering and speed control. RRT* + PD was used to replace the human action of the original human-in-the-loop framework, so that the whole system could run autonomously:

RRT* + PD as expert Human action as expert
Observation LiDAR obstacles, left/right distance to road edge Human eye analyzes the image
Goal Simulator global goal; RRT* local (intermediate) goal Human determines where to navigate
Control action PD controller tracking the local goal Human uses keyboard to steer vehicle
Dynamic model State information is required Not needed
RRT* planning in an ego-centric occupancy grid. The planner searches the drivable region and returns a collision-free local path.
RRT* planning in an ego-centric occupancy grid. The planner searches the drivable region and returns a collision-free local path.

Visual Policy Comparison

Both the learned and expert actions are rolled forward using a kinematic bicycle model. Their predicted trajectories are rendered into the current ego-view, and the VLM periodically compares them based on safety and progress toward the goal.

Current observation transformed into the modified observation fed to the VLM, with the red expert trajectory and blue learner trajectory overlaid.
The current ego-view observation is annotated with predicted red (expert) and blue (learner) trajectories to form the modified observation fed to the VLM.

From Policy Switching to Learning

When the VLM prefers the expert trajectory, the expert action is treated as a local demonstration for behavior cloning. A critic is also trained from the VLM preference signal using TD-style learning, allowing these local comparisons to influence future policy behavior.

Evaluation in MetaDrive

The learned policy improved from near-zero success early in training to a high-success regime, while crashes decreased and reliance on the expert reduced over time. Evaluation was performed over 50 full episodes without intervention.

Metric Value
Success rate 86%
Crash rate 0%
Out-of-road rate 8%
Route completion 93.5%
Animated rollout of the trained VLAPS policy driving in MetaDrive, with predicted trajectory overlays visible in the HUD.
Representative rollout of the trained policy in MetaDrive.
VLAPS training and evaluation trends in MetaDrive: success increases over training while crash rate decreases.
VLAPS training and evaluation trends in MetaDrive. Success increases over training while crash rate decreases.

Practical Limitations

Key Takeaway

VLAPS replaces continuous human supervision with a VLM-based policy gate that decides when an expert should take over and uses the same signal to improve the learned policy.

  • Autonomous expert: RRT* + PD
  • VLM-based policy comparison
  • Behavior cloning from expert-preferred states
  • TD-style value learning
  • No human intervention during training
  • 86% final success with zero crashes in the tested MetaDrive evaluation