End-to-End Safe Social Navigation via Multi-Task Reinforcement Learning and Probabilistic Perception

Tommaso Van Der Meer; Andrea Garulli; Antonio Giannitrapani; Renato Quartullo; Alberto Vaglio; Alexandre Alahi

cs.ROOct 5, 2026

End-to-End Safe Social Navigation via Multi-Task Reinforcement Learning and Probabilistic Perception

Authors: Tommaso Van Der Meer, Andrea Garulli, Antonio Giannitrapani, Renato Quartullo, Alberto Vaglio, Alexandre Alahi

Organizations: Dipartimento di Ingegneria dell’Informazione e Scienze Matematiche, Universit`a di Siena, Siena, Italy · Uninettuno University, Rome, Italy · Visual Intelligence for Transportation (VITA) Lab, ´Ecole Polytechnique F´ed´erale de Lausanne (EPFL), Lausanne, Switzerland

Abstract

Autonomous social navigation requires balancing efficiency, physical safety, and social compliance. Reinforcement Learning (RL) methods provide a viable and effective solution but often rely on unrealistic assumptions, such as the knowledge of humans' position and velocity. In this paper, we introduce JESSI (JAX-based E2E Safe Social Interpretable navigation), a lightweight end-to-end RL framework that maps raw LiDAR scans directly to kinematically feasible control commands. JESSI enhances safety via Dirichlet-parameterized continuous action spaces and deterministic bounding, while an integrated attention-based perception module extracts probabilistic human states for interpretable, socially aware decision-making. Through extensive simulations and real-world deployment on a differential-drive robot, we demonstrate that jointly optimizing the RL policy with a supervised perception signal in a multi-task paradigm enhances social behavior. Ultimately, JESSI is able to balance high navigation success rates and superior social behaviors compared to state-of-the-art baselines.

Figures & tables

Fig. 1: Overview of the proposed method. JESSI operates directly from raw LiDAR scans and leverages a perception module to extract probabilistic human states (human-centered bivariate gaussians) and plan socially-aware and kinematically feasible control commands through a scene attention mechanism and Dirichlet distributions.

Fig. 2: Triangular action space of the differential drive with linear velocity $v\in[0,\bar{v}]$ and angular velocity $\omega\in[-\bar{\omega},\bar{\omega}]$ .

Panel of Figure 3 — Fig. 3: Complete architecture of the JESSI policy. The top part in blue represents the perception module while the bottom part in red depicts the actor-critic module. The inputs of the overall network are depicted in green blocks. In order, the dimensionalities $N_{t}$ , $N_{b}$ , $N_{e}$ , $N_{s}$ , $N_{k}$ represent, the LiDAR stack dimension, the number of LiDAR beams for each stack, the embedding size, the number of angular sectors of the cross-attention layer, and the number of objects predicted by the perception module.

Fig. 4: Example of one scenario in the first experimental setup (parallel traffic) with 5 humans (in blue) and 5 box-shaped obstacles (in black). The robot (circle) and its goal (star) are depicted in red. Humans continuously flow in the corridor in the opposite direction of the robot.

Fig. 5: Example of the second experimental setup with 5 humans (in blue) and 5 circular obstacles (in black). The robot (circle) and its goal (star) are depicted in red. Humans continuously flow in the center of the scene.

Policy	SR (%)	CR-H (%)	CR-O (%)	TR (%)	TtG (s)	LJ (m/s 3 )	AJ (rad/s 3 )	SC (%)
JESSI-MT	94.2	0.0	1.8	4.0	18.25	1.30	4.84	90.1
JESSI-MD	94.9	0.3	1.7	3.1	17.88	1.46	4.88	88.2
JESSI-PO	95.1	0.1	2.2	2.6	18.03	1.77	5.45	87.9
V-E2E	86.3	0.2	12.6	0.9	16.93	0.76	2.27	87.6
BV-E2E	92.9	0.2	1.5	5.4	17.01	1.04	1.91	85.0
DIR-SAFE	73.2	1.1	4.5	21.2	23.17	2.99	10.82	92.1

TABLE I: First experimental setup results

Policy	SR (%)	CR-H (%)	CR-O (%)	TR (%)	TtG (s)	LJ (m/s 3 )	AJ (rad/s 3 )	SC (%)
JESSI-MT	99.9	0.0	0.0	0.1	20.55	0.97	2.96	89.7
JESSI-MD	99.8	0.0	0.0	0.2	19.82	1.11	2.98	87.5
JESSI-PO	100.0	0.0	0.0	0.0	19.73	1.25	3.07	87.9
V-E2E	99.0	0.0	1.0	0.0	18.82	0.65	1.61	87.3
BV-E2E	99.8	0.0	0.2	0.0	17.58	0.83	1.31	80.7
DIR-SAFE	98.1	0.0	0.1	1.8	22.65	2.63	7.52	91.4

TABLE II: Second experimental setup results

Fig. 6: Experimental setup of the real world test with the Loomo Segway robotic platform.

Explore similar work

KinematicRL: A Sim-to-Real Reinforcement Learning Framework For Social Navigation With Kinodynamic Feasibility
Jun 10, 2026Zhiming Xu, Haodong Yang, Chengju Liu +2Robot Navigation Sim-To-Real Reinforcement Learning
School of Computer Science and Technology, Tongji University, Shanghai 201804, China · Department of Electronics and Information Engineering, Tongji University, Shanghai 201804, China · Shanghai Institute of Intelligent Science and Technology, Tongji University, Shanghai 201210, China
Learning Social Robot Navigation By Sensing Human Legs
Jul 30, 2026Alberto Vaglio, Andrea Garulli, Antonio Giannitrapani +2Robot Navigation Two-Dimensional Light Detection And Ranging
Department of Information Engineering and Mathematics, University of Siena, Siena, Italy · Uninettuno University, Rome, Italy
Think When It Matters: Conditional VLM Reasoning for Social Navigation with RL Policies
Jul 13, 2026Ali Ahmadi, Hamed Rahimi, Adrien Jacquet Cretides +3Robot Navigation Safe Navigation
Institut des Syst`emes Intelligents et de Robotique (ISIR) Sorbonne Universit´e France

Jun 10, 2026, cs.ROJ/K move · Enter open · S save

KinematicRL: A Sim-to-Real Reinforcement Learning Framework For Social Navigation With Kinodynamic Feasibility

Zhiming Xu, Haodong Yang, Chengju Liu, Qijun Chen+1

School of Computer Science and Technology, Tongji University, Shanghai 201804, China · Department of Electronics and Information Engineering, Tongji University, Shanghai 201804, China · Shanghai Institute of Intelligent Science and Technology, Tongji University, Shanghai 201210, China

Robot Navigation Sim-To-Real Reinforcement Learning Safe Navigation Two-Dimensional Light Detection And Ranging

Open paper AskPDF arXiv

Deep Reinforcement Learning (DRL) has shown promise for social navigation, yet its real-world deployment remains hindered by a persistent sim-to-real gap arising from simplified first-order dynamics and context-specific human state estimation pipelines. This work presents a unified framework that addresses these limitations to produce dynamically feasible navigation policies suitable for real-world deployment. First, theoretical analysis reveals that tracking error between simulated and actual robot position decays exponentially with increased control order, motivating the use of higher-order control inputs as DRL action space. A second-order control formulation tailored to differential drive robots is developed, complemented by a stochastic iterative Linear Quadratic Regulator (iLQR) that pretrains the policy via a divergence minimization objective. Second, to avoid the added system complexity of camera-LiDAR fusion, a cluster-based human tracking pipeline using only 2D LiDAR is introduced. Human detections are associated according to both spatial proximity and velocity similarity, enabling reliable differentiation of nearby pedestrians and yielding stable velocity estimates through temporal aggregation. Third, we introduce an unbiased residual gating block to balance reaction- and memory-based behaviors while handling time-varying crowd sizes, both critical for social navigation. The resulting policy, KinematicRL, consistently improves kinematic performance and adapts to varying number of detected humans. Experiments in real-world environments demonstrate that, when combined with the proposed tracking pipeline, KinematicRL can be deployed on a real differential drive robot with minimal modifications.

Abstract

Explore similar work

KinematicRL: A Sim-to-Real Reinforcement Learning Framework For Social Navigation With Kinodynamic Feasibility

Learning Social Robot Navigation By Sensing Human Legs

Think When It Matters: Conditional VLM Reasoning for Social Navigation with RL Policies