Contingency-Aware Diffusion Planner (CoPlanner)

• The paper introduces a contingency-aware planner that integrates a pivot-conditioned diffusion mechanism with multi-scenario risk scoring for robust multi-agent trajectories.
• It employs a two-stage process by first generating a safety-validated pivot segment and then branching into diverse, long-term trajectories to ensure interaction-aware, socially consistent planning.
• Empirical evaluations on nuPlan show improved trajectory smoothness, reduced collisions, and enhanced safety in complex driving scenarios.

A contingency-aware diffusion planner (“CoPlanner” as described in (Zhong et al., 21 Sep 2025)) is an integrated generative planning framework that jointly addresses multi-agent interactive trajectory generation and robust decision-making under scenario uncertainty. CoPlanner is designed to address critical limitations in existing generation-then-evaluation motion planning systems, particularly their tendency to adopt a single most likely outcome and their inability to retain fallback trajectories essential for safety in rare but critical scenarios. CoPlanner incorporates a pivot-conditioned diffusion mechanism to ensure consistent, validated immediate behavior, and evaluates candidate trajectories across a spectrum of plausible long-term outcomes using a multi-scenario scoring strategy. This elevates robustness against multi-modal future uncertainty and enables interaction-aware, socially consistent planning for autonomous driving in complex environments.

1. Pivot-Conditioned Diffusion Mechanism

At the heart of CoPlanner is the pivot-conditioned diffusion process that divides planning into two temporally coupled components:

• Short-Term Shared “Pivot” Segment: For all agents, the planner first generates a short-term segment (0 … t_b) that is validated to satisfy safety and traffic rule compliance. This segment forms an “anchor” for all downstream plans.
• Branching Long-Term Futures: Conditional on the selected validated pivot, CoPlanner stochastically generates a set of diverse, long-horizon branches for both the ego-agent and surrounding agents. These diverse futures capture the multi-modality of plausible scene evolutions beyond the immediate past.

Mathematically, the joint distribution over multi-agent trajectories Λ is factorized as:

$$p(\Lambda \mid C) = p_\theta(S \mid C) \, p_\phi(B \mid S, C)$$

where $S$ (the pivot) is the joint short-term segment, $B$ the set of long-term branches, and $C$ the conditioning context.

Sampling employs a temporal mask $m(t) = 1[t \leq t_b]$ during reverse diffusion. At each step:

• For $t \leq t_b$, the known anchor is re-noised (to preserve the pivot).
• For $t > t_b$, the normal denoising process is applied, allowing multimodal branching.

This technique ensures all candidates remain temporally anchored to a safety-validated, interaction-consistent short-term evolution, while retaining diversity over plausible long-term developments.

2. Contingency-Aware Multi-Scenario Scoring

To select robust plans, CoPlanner abandons single-hypothesis selection in favor of multi-scenario evaluation:

• For each candidate short-term anchor, multiple long-term branches are generated via the pivot-conditioned diffusion model.
• Each complete candidate plan is evaluated both on the shared segment and across all generated branches, using a composite score that captures safety, progress, comfort, and rule compliance.

Let $J(\tau^0 \mid \Lambda, C)$ be the cost for ego plan $\tau^0$ under long-term multi-agent realization $\Lambda$ and context $C$, decomposed as:

$$J(\tau^0 \mid \Lambda, C) = J_{shared}(\tau^0_{0:t_b} \mid \Lambda, C) + J_{branch}(\tau^0_{t_b:T} \mid \Lambda, C)$$

For each candidate, the risk-aggregated score across $N$ branches is:

$$J_k(\tau^0) = J_{shared}(\cdot) + R(\{J_{branch}(\cdot)\}_n)$$

where $R$ is a risk aggregator (e.g., equal-weight mean). The plan minimizing this overall score is selected as the robust, contingency-aware trajectory.

3. Technical Architecture and Diffusion Model

CoPlanner employs a joint generative model for multi-agent trajectories in a discrete-time, variance-preserving diffusion process. The model backbone is a diffusion transformer (DiT), which enables joint modeling of all dynamic agents (ego plus surroundings).

The training objective is an $x^{(0)}$-prediction loss under the VP schedule:

$$L_\theta = \mathbb{E}_{x^{(0)}, h, x^{(h)}} \left\| \mu_\theta(x^{(h)}, h, C) - x^{(0)} \right\|_2^2$$

Reverse diffusion is adapted to accept the pivot-masked temporal conditioning. The model ensures that the trajectory generation for all agents is coupled during the anchor segment and allowed to diversify beyond.

4. Integration of Prediction and Planning Under Interaction

Crucially, CoPlanner integrates multi-agent prediction and ego motion planning into a unified generative framework:

• All joint trajectories (ego + agents) are sampled together, ensuring social consistency (i.e., the planned ego action does not conflict with the simultaneously forecasted neighbor actions).
• The pivot segment validation guarantees that near-term execution is robust—no plan can “drift away” into unsafe or implausible behavior immediately after re-planning, a critical requirement in highly interactive, uncertain road scenes.
• The diverse branch generation preserves fallback options: if the realized future diverges from the likely prediction, the system can rely on pre-evaluated alternative continuations.

Unlike conventional decoupled architectures—where planning passively reacts to exogenous predictions—CoPlanner’s joint modeling and evaluation actively anticipates interactive evolution across the ensemble of sampled futures.

5. Empirical Performance on nuPlan and Safety Outcomes

CoPlanner was benchmarked in closed-loop (planning-in-the-loop) evaluations on the nuPlan driving dataset. Performance results highlight the utility of the contingency-aware, pivot-conditioned approach:

• On Val14 and Test14, as well as Test14-hard subsets, CoPlanner achieved higher overall scores across both non-reactive (log-replay) and reactive (IDM controller) settings than state-of-the-art baselines (Diffusion Planner, UrbanDriver, PDM).
• Notable improvements were observed in:
  • Fewer collisions and increased time-to-collision (TTC)
  • Greater trajectory smoothness and driver comfort metrics
  • Robust operation in challenging scenarios involving rare and adversarial behaviors (due to fallback-preserving branch structure)
• The two-stage sampling and risk-aggregated evaluation directly contributed to these robustness gains across both familiar and previously unseen test conditions.

6. Practical Implications and Extensions

CoPlanner’s architecture provides several practical advantages for safety-critical real-world deployment:

• Preservation of fallback options: The ensemble of plausible long-term continuations ensures that if traffic evolves unexpectedly, alternative, pre-validated strategies remain available.
• Social consistency: By jointly generating multi-agent scenarios, the system avoids common conflicts and unrealistic downstream plans introduced by exogenously predicted agent behavior.
• Flexible risk aggregation: The scoring framework supports diverse risk measures and could, with extension, incorporate distributional uncertainty, learned risk attitudes, or dynamically chosen scenario weights.
• Compatibility with physical constraints: The design can accommodate future improvements—such as incorporating feasibility-aware guidance, physics-informed priors, or adaptive branching-points (t_b)—to further align generated plans with vehicle dynamics and scene semantics.

Potential future directions include adaptively setting the short-term branching time per situation, learning scenario weights for better uncertainty calibration, and broader closed-loop evaluation with diverse and more adversarial reactive agent sets.

7. Mathematical Formulation Table

Component	Mathematical Expression	Description
Short/Long Decomposition	$$p(\Lambda | C) = p_\theta(S | C) \, p_\phi(B | S, C)$$	Joint future as anchor + branches
Risk-Aggregated Cost	$$J_k(\tau^0) = J_{shared}(\cdot) + R(\{J_{branch}(\cdot)\}_n)$$	Selection over multi-scenario cost
Diffusion Loss	$$L_\theta = \mathbb{E}_{x^{(0)}, h, x^{(h)}} \left\| \mu_\theta(x^{(h)}, h, C) - x^{(0)} \right\|_2^2$$	Model objective under VP schedule
Pivot Masking	$m(t) = 1[t \leq t_b]$	Time mask for anchor/branch split

These formalizations anchor CoPlanner’s technical contributions and establish its foundation as a contingency-aware, interaction-consistent, and robust framework for autonomous motion planning.

---

CoPlanner’s interactive, scenario-robust design represents a major advancement toward closing the gap between offline generative modeling and online, safety-critical execution in autonomous driving systems (Zhong et al., 21 Sep 2025).