arXiv subjects

Rodrigue de Schaetzen

Publications and source records attributed to Rodrigue de Schaetzen.

7 recordsLinked to original sources

arXiv2026

Constrained Group Relative Policy Optimization

Group Relative Policy Optimization (GRPO) remains the dominant critic-free approach for fine-tuning LLMs and VLMs, but its compatibility with constrained policy optimization (e.g. for safety-critical domains) has not been carefully examined. In this work, we introduce Constrained GRPO, a Lagrangian-based extension of GRPO for constrained policy optimization. We show that the standard practice of scalarizing rewards before normalization introduces a critical Lagrangian-specific failure mode: GRPO's within-group normalization makes constrained optimization highly sensitive to how multi-component learning signals are aggregated. We show that scalarizing rewards before normalization introduces shared-denominator coupling, so that changing one multiplier alters not only the emphasis on its corresponding constraint, but also the relative weighting of the reward and other constraints. We address this with a simple but crucial modification: scalarizing standardized advantages rather than rewards. This yields a better-conditioned update by addressing the coupling induced by reward scalarization, resulting in better-behaved multiplier dynamics and more stable constraint enforcement in practice. Empirically, across a controlled gridworld, a real-world autonomous driving benchmark, and a mathematical reasoning task, Constrained GRPO consistently achieves better adherence to specified constraints while maintaining or improving task performance.

cs.LG↗

arXiv2026

Scaling Self-Play for End-to-End Driving

End-to-end autonomous driving models are typically trained on offline human-demonstration datasets that provide limited state coverage and often no closed-loop feedback, making them prone to compounding errors when deployed in closed-loop and brittle to long-tail agent interactions. To overcome these limitations, we propose an alternative strategy for training end-to-end driving models: large-scale self-play directly from pixels in simulation. While prior self-play approaches have shown promising transfer to real-world driving, they typically assume vectorized Bird's-Eye-View (BEV) observations that are incompatible with end-to-end policies operating directly on sensor observations. To this end, we introduce Gigapixel, a high-throughput batched driving simulator with perspective rendering, enabling scalable self-play directly from pixel observations. Rather than targeting compute-costly photorealistic sensor simulation, Gigapixel renders a simplified bounding-box world that preserves essential scene structure while achieving throughput at 50k agent steps per second. Since direct pixel-space self-play RL is prohibitively sample-inefficient at end-to-end model scale, we propose self-play DAgger training: we train pixel-based policies in self-play via on-policy distillation from a privileged RL teacher. To bridge the sim-to-real gap, we subsequently transfer the self-play trained policies to real-world sensor data through lightweight perception adaptation. Policies trained in Gigapixel and adapted to real-world sensor data achieve competitive performance on the HUGSIM and NAVSIM-v2 benchmarks without human trajectory supervision. Moreover, scaling self-play training yields proportional gains in policy performance, establishing self-play as a practical and scalable strategy for training end-to-end models.

cs.RO↗

arXiv2025

Poutine: Vision-Language-Trajectory Pre-Training and Reinforcement Learning Post-Training Enable Robust End-to-End Autonomous Driving

Maintaining good driving behavior in out-of-distribution scenarios remains a critical challenge in autonomous driving. A promising direction is to leverage the generalist knowledge and reasoning capabilities of large-language models by treating unusual driving scenarios as a logical reasoning task. In this work, we present Poutine, a method that uses an off-the-shelf 3B-parameter vision-language model (VLM) - without any additional components - to achieve robust end-to-end autonomous driving via a simple and scalable training recipe. To learn strong base driving capabilities, we first train Poutine-Base using self-supervised next-token prediction over vision, language, and trajectory (VLT) tokens, leveraging both nominal and long-tail driving data. In the second stage, we fine-tune Poutine-Base using Group Relative Policy Optimization (GRPO) with a small set of human preference-labeled examples. We evaluated our approach on the Waymo end-to-end driving benchmark curated for long-tail scenarios. The final Poutine model achieves an RFS of 7.99 on the test set, placing 1st in the 2025 Waymo Vision-Based End-to-End Driving Challenge by a significant margin. Our results suggest that handcrafted tokenizers or custom architectural components added to base VLMs in prior work are not necessary to achieve strong driving performance. Instead, this work highlights the potential of scalable VLT pretraining combined with lightweight RL fine-tuning to enable robust and generalizable autonomous driving.

cs.RO↗

arXiv2025

AUTO-IceNav: A Local Navigation Strategy for Autonomous Surface Ships in Broken Ice Fields

Ice conditions often require ships to reduce speed and deviate from their main course to avoid damage to the ship. In addition, broken ice fields are becoming the dominant ice conditions encountered in the Arctic, where the effects of collisions with ice are highly dependent on where contact occurs and on the particular features of the ice floes. In this paper, we present AUTO-IceNav, a framework for the autonomous navigation of ships operating in ice floe fields. Trajectories are computed in a receding-horizon manner, where we frequently replan given updated ice field data. During a planning step, we assume a nominal speed that is safe with respect to the current ice conditions, and compute a reference path. We formulate a novel cost function that minimizes the kinetic energy loss of the ship from ship-ice collisions and incorporate this cost as part of our lattice-based path planner. The solution computed by the lattice planning stage is then used as an initial guess in our proposed optimization-based improvement step, producing a locally optimal path. Extensive experiments were conducted both in simulation and in a physical testbed to validate our approach.

cs.RO↗

arXiv2023

Efficient Ground Vehicle Path Following in Game AI

This short paper presents an efficient path following solution for ground vehicles tailored to game AI. Our focus is on adapting established techniques to design simple solutions with parameters that are easily tunable for an efficient benchmark path follower. Our solution pays particular attention to computing a target speed which uses quadratic Bezier curves to estimate the path curvature. The performance of the proposed path follower is evaluated through a variety of test scenarios in a first-person shooter game, demonstrating its effectiveness and robustness in handling different types of paths and vehicles. We achieved a 70% decrease in the total number of stuck events compared to an existing path following solution.

cs.AI↗

arXiv2023

Real-Time Navigation for Autonomous Surface Vehicles In Ice-Covered Waters

Vessel transit in ice-covered waters poses unique challenges in safe and efficient motion planning. When the concentration of ice is high, it may not be possible to find collision-free trajectories. Instead, ice can be pushed out of the way if it is small or if contact occurs near the edge of the ice. In this work, we propose a real-time navigation framework that minimizes collisions with ice and distance travelled by the vessel. We exploit a lattice-based planner with a cost that captures the ship interaction with ice. To address the dynamic nature of the environment, we plan motion in a receding horizon manner based on updated vessel and ice state information. Further, we present a novel planning heuristic for evaluating the cost-to-go, which is applicable to navigation in a channel without a fixed goal location. The performance of our planner is evaluated across several levels of ice concentration both in simulated and in real-world experiments.

cs.RO↗

arXiv2022

Convolutional neural network with a hybrid loss function for fully automated segmentation of lymphoma lesions in FDG PET images

Segmentation of lymphoma lesions is challenging due to their varied sizes and locations in whole-body PET scans. This work presents a fully-automated segmentation technique using a multi-center dataset of diffuse large B-cell lymphoma (DLBCL) with heterogeneous characteristics. We utilized a dataset of [18F]FDG-PET scans (n=194) from two different imaging centers, including cases with primary mediastinal large B-cell lymphoma (PMBCL) (n=104). Automated brain and bladder removal approaches were utilized as preprocessing steps to tackle false positives caused by normal hypermetabolic uptake in these organs. Our segmentation model is a convolutional neural network (CNN) based on a 3D U-Net architecture that includes squeeze and excitation (SE) modules. Hybrid distribution, region, and boundary-based losses (Unified Focal and Mumford-Shah (MS)) were utilized that showed the best performance compared to other combinations (p<0.05). Cross-validation between different centers, DLBCL and PMBCL cases, and three random splits were applied on train/validation data. The ensemble of these six models achieved a Dice similarity coefficient (DSC) of 0.77 +- 0.08 and Hausdorff distance (HD) of 16.5 +-12.5. Our 3D U-net model with SE modules for segmentation with hybrid loss performed significantly better (p<0.05) as compared to the 3D U-Net (without SE modules) using the same loss function (Unified Focal and MS loss) (DSC= 0.64 +-0.21 and HD= 26.3 +- 18.7). Our model can facilitate a fully automated quantification pipeline in a multi-center context that opens the possibility for routine reporting of total metabolic tumor volume (TMTV) and other metrics shown useful for the management of lymphoma.

physics.med-ph↗