Active Interaction-Aware Model Predictive Path Integral via Ego-Conditioned Generative Predictions

Technical Summary: Active Interaction-Aware Model Predictive Path Integral via Ego-Conditioned Generative Predictions

Problem Statement
Autonomous navigation in dense traffic is fundamentally an interactive problem where the ego vehicle and surrounding agents continuously influence one another. Traditional "predict-then-plan" approaches decouple these processes, predicting agent trajectories independently of the ego's future actions. This paradigm suffers from the "frozen robot" problem, where the system becomes overly cautious or deadlocked in interactive scenarios (e.g., merging, unprotected turns) because it fails to account for how its own actions provoke reactions from others. While game-theoretic methods attempt to model these interactions, they often rely on unrealistic assumptions of rationality and suffer from computational intractability in dense traffic. Conversely, existing ego-conditioned planners often adopt a rigid "leader-follower" structure that can lead to overly aggressive behavior by implicitly assuming other agents will yield. The core challenge is to develop a planning framework that actively reasons about interaction uncertainty, probes the hidden intent of other agents, and balances task efficiency with safety without relying on simplified parametric models.

Methodology
The authors propose a hybrid planning framework that integrates a Generative Autoregressive Prediction Model (GARPM)—specifically a variant of the SMART transformer architecture—within a Model Predictive Path Integral (MPPI) control scheme.

1. Ego-Conditioned Prediction: Unlike standard predictors, the GARPM generates stochastic, multi-modal predictions of surrounding agents' futures conditioned on specific candidate ego trajectories. The joint distribution of agent futures is factorized autoregressively, where the next state of every agent is predicted based on the history of all agents and the specific ego trajectory being evaluated.
2. Nested Sampling Scheme: To resolve the circular dependency between planning and prediction (where the plan requires a prediction, but the prediction requires a plan), the authors employ a tractable nested sampling approach:
   • Outer Loop (MPPI): The planner samples $M$ candidate ego trajectories by perturbing a nominal control sequence.
   • Inner Loop (Behavior Rollouts): For each sampled ego trajectory, the GARPM generates $B$ stochastic behavior rollouts for surrounding agents.
   • Surrogate Distribution: To make cost evaluation tractable, the discrete set of $B$ rollouts is converted into a continuous Gaussian Mixture (MoG) surrogate. The means of the Gaussians correspond to the sampled agent states, while the mixing weights are derived from the autoregressive likelihoods of the sampled trajectories.
3. Risk-Aware Cost Evaluation: The stage cost for each ego trajectory is computed by evaluating tracking and speed objectives against the MoG surrogate. Crucially, the collision risk is estimated by integrating the MoG distribution over the collision region. This allows the planner to assess the probability of collision under the induced distribution of agent behaviors.
4. Active Uncertainty Reduction: The framework implicitly incentivizes "active" behavior. Trajectories that actively probe the environment (e.g., inching forward to elicit a reaction) tend to disambiguate agent intents, concentrating the MoG weights on fewer, more consistent hypotheses. This reduces the estimated collision risk in the cost function, naturally guiding the MPPI optimizer toward maneuvers that resolve uncertainty.

Key Contributions
The paper makes three primary contributions:

1. Hybrid Framework: A novel integration of learned autoregressive generative models into MPPI, closing the loop between prediction and planning. This leverages the expressive capacity of generative models to capture multi-modal uncertainty from real-world data while retaining the safety-relevant structure of model-based control.
2. Tractable Interaction-Aware Formulation: An interaction- and uncertainty-aware planning formulation using a nested sampling scheme. This approach evaluates multi-modal predictions conditioned on sampled ego trajectories, yielding risk-aware collision probability estimates without the "leader-follower" bias of standard ego-conditioned planners.
3. Implicit Active Probing: Demonstration that combining generative models with MPPI weighting mechanisms enables active uncertainty reduction. The system achieves more assertive and efficient driving without requiring explicit belief-space representations or dedicated exploration terms.

Results
The proposed method was evaluated in the nuPlan simulator across three interactive urban scenarios: adjacent lane merging, on-ramp merging, and unprotected left turns. The system was compared against four baselines:

• Baseline 1: Multi-modal Predict-Then-Plan.
• Baseline 2: Receding-Horizon Ego-Conditioning (conditioned only on the previous optimal plan).
• Baseline 3: Unimodal Ego-Conditioned Planning.
• Baseline 4: Passive Ego-Conditioned MPPI (fixed probabilities).

Quantitative Findings:

• Merging: The proposed method achieved a 75% merge success rate with the lowest merge time (11.8s) compared to all baselines. Baseline 1 (predict-then-plan) had a 37.5% success rate due to excessive conservatism.
• On-Ramp Merging: The proposed method achieved 0% collision rate, whereas ego-conditioned baselines (2, 3, and 4) suffered collision rates between 15% and 25% due to overestimating cooperativeness. Baseline 1 remained overly conservative.
• Unprotected Left Turn: The proposed method achieved the lowest deadlock rate (5%) and collision rate (0%), outperforming baselines which exhibited deadlock rates up to 30% and collision rates up to 10%.

Qualitative Findings:
Visual analysis (Figs. 2–6) confirms that the proposed method successfully infers the cooperativeness of surrounding agents through active probing. In cooperative scenarios, the planner commits to the maneuver once the "yield" mode becomes dominant. In non-cooperative scenarios, the "proceed" mode dominates, increasing the estimated collision risk and causing the planner to select safer, evasive maneuvers. In contrast, baselines either failed to resolve intent ambiguity (leading to deadlock) or remained overly optimistic (leading to collisions).

Significance
The paper claims that its significance lies in providing a principled solution to the "frozen robot" problem without the computational burden of game-theoretic equilibrium computation or the behavioral limitations of simplified parametric models. By leveraging the autoregressive nature of modern generative models within a sampling-based controller, the framework allows autonomous vehicles to actively probe their environment to reduce uncertainty. This results in driving behavior that is simultaneously safer, more efficient, and more decisive than conventional predict-then-plan or passive interaction-aware approaches, effectively mimicking the active uncertainty reduction strategies observed in human drivers.