Before Parc Fermé: RL-Time Pruning for Efficient Embodied LLMs in Autonomous Driving

Abstract: Embodied LLMs are increasingly used as reasoning modules in robotic control pipelines to improve human-robot interaction, but their memory and generation latency make real-time deployment difficult. Pruning can reduce these costs, but for controllers that undergo multiple pre- and post-training phases, the crucial question is not only how much to prune, but when pruning should occur. In this work, we propose Before Parc Fermé (BPF), a pruning strategy performed during RL that compresses embodied LLM controllers while they are still being optimized for closed-loop behavior. This allows pruning decisions to account for the task-specific supervision and closed-loop feedback that shape the final controller. We propose two variants: BPF-RL, which performs iterative pruning during RL by removing part of the model at predefined training intervals, and BPF-SFT/RL, which first prunes part of the model structure during SFT and then further compresses it during RL using the same iterative strategy as BPF-RL until the target pruning ratio is reached. We evaluate BPF on RobotxR1, an LLM-based autonomous-driving control pipeline, using an established LLM pruning framework (LLM-Pruner), and compare it against post-training pruning, post-training pruning with RL recovery, SFT-stage pruning, and smaller dense models from the same family. Our results show that BPF provides the best task-performance vs. memory and throughput trade-off among the considered pruning strategies. When compressing the larger RobotxR1 models, BPF-SFT/RL achieves a $1.69\times$ better size-end-to-end performance trade-off than directly selecting a smaller dense model from the same family, measured as removed parameters per lost percentage point of control adaptability. On the Jetson AGX Orin mounted on the target robotic platform, the compact models improve decode throughput by up to $27\%$.

• The paper introduces BPF-SFT/RL, which combines Taylor-based structured pruning with ongoing GRPO optimization and outperforms post-training pruning, achieving a 1.69× better compression–adaptability trade-off.
• BPF-SFT/RL preserves strong driving performance at high compression, with 45% pruning improving adaptability by 17% over post-training pruning and 10% pruning matching the strongest unpruned baseline while removing 0.25B parameters.
• Deployment tests show that 45% DecisionxR1 pruning cuts memory from 2.07 GB to 1.51 GB and raises throughput by 27%, but end-to-end latency also depends on generated-token length and module interactions.

Motivation and problem statement

Embodied LLM controllers trained with SFT followed by closed-loop RL, such as the RobotxR1 autonomous-driving pipeline, face a deployment dilemma: larger models learn more robust control behavior but exceed the memory and latency budgets of onboard hardware. In RobotxR1, moving from Qwen2.5-1.5B to Qwen2.5-3B improves control adaptability by more than 20% but increases token-generation throughput by over 60%. Structured pruning is an obvious remedy, but the paper's central observation is that prior work treats pruning as a post-training operation, ignoring when compression should occur in a multi-phase training pipeline. If pruning happens after SFT but before RL, the pruning criterion lacks access to closed-loop feedback; if it happens after full training, remaining weights have little opportunity to adapt to removed structures.

The paper proposes Before Parc Fermé (BPF), a family of pruning schedules that compresses embodied LLM controllers while they are still being optimized for closed-loop behavior — before the "parc fermé" stage in which the architecture is frozen. Two variants are studied: BPF-RL, which performs iterative structured pruning during RL, and BPF-SFT/RL, which applies partial pruning during SFT and completes the target ratio iteratively during RL.

Methodology

BPF is instantiated on the RobotxR1 pipeline, which comprises two sequentially executed RAG-enhanced LLM modules built on Qwen2.5: DecisionxR1, which judges whether current vehicle behavior matches the user instruction, and MPCxR1, which generates textual updates to exposed Model Predictive Control parameters. Both modules are adapted via SFT followed by GRPO-based RL, using behavior-adherence rewards (DecisionxR1) and closed-loop driving-performance rewards (MPCxR1). The controller, prompts, rewards, and evaluation protocol are inherited unchanged from RobotxR1.

The pruning criterion is the first-order Taylor importance score from LLM-Pruner, computed on calibration batches at each pruning event:

$$I(w_i) = \left| w_i \frac{\partial \mathcal{L}(\mathcal{D};\theta)}{\partial w_i} \right|$$

with scores aggregated over coupled parameter groups to preserve architectural consistency. Structured pruning is chosen because it yields smaller dense architectures compatible with standard inference runtimes; unstructured results appear only as a module-level robustness check in the appendix.

The compared schedules are:

• Post-training pruning (PTP): one-shot pruning after full SFT-RL.
• PTP+R: PTP followed by additional RL steps for recovery.
• SFT-stage pruning: single-shot pruning mid-SFT.
• BPF-RL: after SFT, a 400-step RL warm-up, then three pruning events spaced 100 steps apart, each removing $\rho/J_{\mathrm{RL}}$ of parameters, with RL resuming between events.
• BPF-SFT/RL: half the target ratio pruned mid-SFT, the remainder spread across the same iterative RL schedule.

All schedules use identical total SFT (200 steps) and RL (900 steps) budgets.

Results on DecisionxR1 pruning

With MPCxR1 fixed, BPF schedules dominate all baselines on the size–control-adaptability Pareto front starting from Qwen2.5-3B. At the 45% pruning ratio, BPF-SFT/RL improves average control adaptability by 11.4% over PTP+R and 17% over PTP; BPF-RL improves by 9.7% and 15.2%, respectively. Notably, SFT-stage pruning fails the task entirely even at the smallest tested ratio, indicating that pruning before any closed-loop feedback is available is not viable in this setting.

Two quantitative claims stand out. First, the 45%-pruned BPF-SFT/RL model achieves a $1.69\times$ better compression–adaptability trade-off than directly selecting the dense Qwen2.5-1.5B model, measured as removed parameters per lost percentage point of control adaptability. Second, the 10%-pruned BPF-SFT/RL model matches the strongest unpruned baseline while shedding 0.25B parameters. Appendix results show that the 1.5B checkpoint has far less pruning headroom: its adaptability drops sharply already at 20% pruning, supporting the thesis that compression should start from a larger backbone.

Joint pipeline pruning

When both modules are pruned, the cascade exhibits strong module asymmetry. Increasing MPCxR1 pruning from 10% to 30% causes substantially larger degradation than equivalent DecisionxR1 pruning — the 30/10 configuration stays within 5% of the unpruned baseline while 30/30 falls to roughly the mid-40% range of average improvement. This is consistent with MPCxR1's role as the direct generator of controller parameters. The best configurations keep MPCxR1 at only 10% pruning while allowing up to 30–45% compression of DecisionxR1. Interestingly, mild joint pruning (the 10/10 configuration) slightly exceeds the unpruned 3B baseline, suggesting BPF removes structures that do not benefit the closed-loop task.

Embedded deployment

Deployment uses Q5_K_M quantization, GGUF export, and llama.cpp on the Jetson AGX Orin mounted on the 1:10-scale racecar, with MLP weights zero-padded to multiples of 256 to avoid FP16 fallbacks. Pruning DecisionxR1 by 45% reduces memory from 2.07 GB to 1.51 GB and raises decode throughput from 45 to 58 tok/s (+27%); MPCxR1 at 30% reaches 55 tok/s at 1.62 GB.

A key finding is that parameter reduction does not guarantee end-to-end latency reduction. Because DecisionxR1 output conditions MPCxR1 generation, pruning can alter generated-token counts — e.g., making MPCxR1 more verbose — offsetting throughput gains. The best latency configuration (DecisionxR1 pruned 20%, MPCxR1 unpruned) reduces pipeline latency by 19.5% with only a 0.8% task-performance drop, and achieves 21.0% higher task performance than the dense 1.5B baseline despite being $1.57\times$ slower. This motivates evaluating embodied LLM compression end-to-end rather than through task-agnostic metrics alone.

Unstructured-pruning appendix results reinforce the scheduling claim: PTP and PTP+R fall below the dense 1.5B standalone accuracy, whereas BPF-SFT/RL retains 85.40% accuracy at 1.85B parameters versus 82.83% for the 1.5B baseline.

Limitations

The authors identify four limitations. First, modules are adapted and pruned separately — MPCxR1 trains on outputs from the unpruned DecisionxR1 — which may explain the verbosity and latency anomalies of some pruned cascades; joint or alternating training of compressed modules remains open. Second, pruning schedules use fixed intervals and increments; adaptive or finer-grained RL-time policies are unexplored. Third, the procedure is not quantization-aware, requiring dimension padding for runtime alignment. Fourth, the RL objective does not penalize generated-token length, even though token count demonstrably governs end-to-end latency.

Conclusion

This paper demonstrates that the training-stage placement of pruning materially affects the quality of compressed embodied LLM controllers. Coupling structured Taylor-based pruning with ongoing GRPO optimization (BPF-SFT/RL) outperforms post-training pruning, post-training recovery, and SFT-stage pruning, and dominates the alternative of deploying a smaller dense model, by $1.69\times$ on the compression–adaptability trade-off. The deployment study adds a practically important caveat: in modular LLM pipelines, compression must be judged by full-cascade latency and task performance, since pruned models can generate longer outputs that negate their throughput advantage.