CarPlanner: Autonomous Vehicle Planning

• CarPlanner is a framework that combines reinforcement learning, imitation learning, and rule-based optimization for robust vehicle trajectory, route, and assignment planning.
• Empirical evaluations show that its configurations achieve state-of-the-art performance on benchmarks like nuPlan, CARLA, and Waymax, ensuring high safety, progress, and comfort.
• The system extends to urban mobility applications such as large-scale parking assignment and multimodal carpool–transit integration using methods like the Hungarian algorithm and BEV-based planning.

CarPlanner refers to a family of algorithmic and system architectures for vehicle trajectory, route, and assignment planning in large-scale, real-world autonomous driving and smart mobility settings. A salient feature of CarPlanner is its coupling of scalable optimization, data-driven learning (including reinforcement learning and imitation learning), and real-time, multi-agent coordination, operationalized in urban navigation, parking assignment, carpool–transit integration, and end-to-end motion planning applications. The term has been used explicitly for advanced reinforcement learning-based planners, as well as paradigmatically in assignment and multi-modal planning literature.

1. Consistent Auto-regressive Trajectory Planning for RL

CarPlanner is instantiated as a Consistent Auto-regressive Planner for large-scale reinforcement learning (RL) in autonomous vehicle trajectory planning (Zhang et al., 27 Feb 2025). The approach seeks to overcome inefficiencies and instability typical in RL-based planners for real-world scenarios.

Problem Formulation

Given the multi-step Markov decision process (MDP) structure, the ego vehicle’s future state sequence is generated using an auto-regressive policy: $$P(s_{0:T}) = \rho_0(s_0)\int P(c|s_0)\prod_{t=0}^{T-1}[\pi(a_t|s_t,c)\cdot P(s_{t+1}^{1:N}|s_t)]\, dc$$ where $c$ encodes time-persistent trajectory mode (longitudinal and lateral), and $P(s_{t+1}^{1:N}|s_t)$ is a world-model predictor for other traffic agents.

Consistency Mechanism

Crucially, the mode $c$ (discrete combinations of lateral/longitudinal behaviors) is fixed per rollout, ensuring temporal consistency and thus reducing policy-induced trajectory incoherence. The entire sequence for a given rollout is conditioned on the fixed $c$.

Generation–Selection Framework

A multi-modal set of trajectories is sampled per mode $c$; each is scored by a rule-augmented selector ($\text{ModeScore}(c) + 0.3\cdot\text{RuleScore}$), optimizing for safety (collision, drivable area), comfort, and progress. The best-scoring trajectory is used, operationalized by an LQR controller.

A universal, expert-guided reward ($R_t = -\text{DE}(s_t^0, s_t^{0,gt}) + Q(s_t)$) enables efficient policy update via PPO without scenario-specific tuning.

Invariant-View Module

A preprocessing module ensures input is time-agnostic and representation-invariant by ego-frame coordinate transformation and KNN-based element selection.

Empirical Results: CarPlanner achieves SOTA closed-loop performance on nuPlan, exceeding both RL, IL, and rule-based baselines: CLS-NR 94.07% (Test14-Random), S-CR 96.38%, S-PR 95.37%, with robust performance on safety, progress, and comfort metrics (Zhang et al., 27 Feb 2025).

2. Adaptive IL-Based Motion Planning with Dynamic Scene Awareness

Another prominent CarPlanner instantiation is CarPLAN, a context-adaptive imitation learning (IL) framework, emphasizing context-awareness via dynamic scene encoding and flexible trajectory decoding (Yun et al., 13 Mar 2026).

Displacement-Aware Predictive Encoding (DPE)

The encoder receives agent state histories, vectorized map data, and goals, predicting future displacement vectors (relative positions) between the AV and all scene elements. DPE is supervised with an auxiliary loss: $$L_{\text{disp}} = \sum_{a=1}^{N_a+N_p}\sum_{t=1}^{T_f} \operatorname{SmoothL1}(D_a^t, \hat D_a^t)$$ forcing extra spatial relational modeling without inference time cost.

Context-Adaptive Multi-Expert Decoder (CMD)

A Mixture-of-Experts (MoE) decoder architecture with a scene-aware router dynamically dispatches trajectory queries to context-relevant experts per traffic scene, enabling specialization across rare or challenging contexts. Load-balancing regularization enforces diverse expert utilization.

Key Empirical Results: CarPLAN achieves SOTA on nuPlan Val14/CLS-NR 91.4, Test14-Hard/CLS-NR 78.9, and generalizes successfully to Waymax, outperforming prior diffusion and transformer-based IL planners (Yun et al., 13 Mar 2026).

3. Parking Assignment and Automated Parking Planning

CarPlanner also denotes scalable methods for parking resource allocation and automated parking maneuver planning.

Large-Scale Parking Assignment

A real-time assignment algorithm solves the assignment of $M$ vehicles to $c$0 ($c$1) parking slots via a two-stage reduction and exact minimization:

• For each vehicle, select $c$2-nearest slot candidates; let $c$3 be their union ($c$4).
• Inflate to square cost matrix with dummy assignments; solve with Hungarian algorithm in $c$5 time. System architecture includes a sensing layer (bay occupancy), communications layer (apps incl. CarPlanner), computation layer (assignment engine), and guidance layer (assignment + routing). In practical city-scale simulations ($c$6, $c$7), batch assignment is near-optimal ($c$8 waste for small $c$9) and responses are sub-5s for moderate batch sizes (Zhao et al., 2017).

Automated BEV-Based Parking

The AVP module integrates BEV vision occupancy, a kinematic-bicycle model, enhanced A*-based global search (with bidirectional search and volume pruning), and final Bezier/B-spline smoothing with NLP-based trajectory optimization: $P(s_{t+1}^{1:N}|s_t)$0 subject to kinematic, bound, and dualized collision constraints. Execution times for typical scenarios are sub-1s, with empirically measured $P(s_{t+1}^{1:N}|s_t)$1+ reductions in acceleration and steering peaks vs. baseline (Zhao, 2024).

4. Multi-Modal Trajectory and Motion Planning with LLMs

PlanAgent, positioned as "CarPlanner," leverages a multi-modal LLM as a reasoning engine for mid-to-mid vehicle planning (Zheng et al., 2024). The system translates BEV maps and text-based lane graphs into structured planning instructions and code via chain-of-thought reasoning, closed-loop simulation reflection, and explicit adherence to traffic rules. The controller parameters the IDM as: $P(s_{t+1}^{1:N}|s_t)$2 where scene and instruction flow is parsed and validated by a reflection loop (short-term safety simulation). PlanAgent surpasses SOTA planner scores on nuPlan (NR-CLS 93.26 on Val14, 72.51 on Test14-hard), and ablations confirm criticality of BEV, text, and reflection for generalization and safety.

5. Integration of CarPlanner with Urban Mobility Systems

CarPlanner schemes also encompass multimodal journey optimization and carpool–public transit integration:

• Meeting Point Recommendations: GIS-prefiltered candidate sets, time-dependent Dijkstra precomputations, and real-time scoring (range-vote sum/minimax) yield optimal pick-up points for ride-sharing, achieving $P(s_{t+1}^{1:N}|s_t)$310ms per-query latency at city scale (Czioska et al., 2017).
• Carpool–Transit Synchronization: CarPlanner models each driver detour as an ephemeral GTFS bus line (PoolLine), injects these into the GTFS feed, and utilizes off-the-shelf planners (e.g. OTP) for multimodal routing. In urban scale studies, this results in a 10 percentage point improvement in rider matching and significant CO₂ savings vs. stand-alone PT or carpooling (Chaabouni et al., 2022).

6. Experimental Performance, Benchmarks, and Limitations

Across RL/IL-based, assignment, and heuristic planners, CarPlanner approaches have defined or surpassed SOTA on public benchmarks (nuPlan, CARLA, Waymax). Strengths include:

• Consistency and temporal coherence (via mode conditioning) in RL (Zhang et al., 27 Feb 2025)
• Specialization to rare context via MoE for IL (Yun et al., 13 Mar 2026)
• Real-time, provably near-optimal parking assignment for arbitrary $P(s_{t+1}^{1:N}|s_t)$4 (Zhao et al., 2017)
• Sub-second integration with large-scale route planning under bans and parking constraints (Kleff et al., 2020)

Documented limitations concern non-reactive agent prediction in RL, parameter sensitivity in MoE load-balancing, and batch-size selection trade-offs in large-scale assignment.

7. Outlook and Research Directions

Future research in the CarPlanner direction targets truly interactive, multi-agent world-modeling, hierarchical RL and IL for horizon expansion, spatial indexing and predictive candidate pruning in assignment problems, machine-learning-backed cost functions in planning, and deeper LLM integration for semantic instruction fusion. The unified CarPlanner framework thus synthesizes advances in RL/IL algorithms, combinatorial optimization, and multimodal integration to advance the field of autonomous and cooperative mobility planning.