DODGER: Safety-Guided Reinforcement Learning for Robot Navigation Among Dynamic Obstacles

Sanghyuk Park; Kwanwoo Lee; Taekyung Kim; Seohyeon Lim; Yisoo Lee

cs.ROSep 30, 2026

DODGER: Safety-Guided Reinforcement Learning for Robot Navigation Among Dynamic Obstacles

Authors: Sanghyuk Park, Kwanwoo Lee, Taekyung Kim, Seohyeon Lim, Yisoo Lee

Organizations: Department of Intelligence and Information, Seoul National University, Republic of Korea · Department of Robotics, University of Michigan, Ann Arbor, MI, USA · Department of Mechanical Engineering, Yonsei University, Republic of Korea · Center for Humanoid Research, Korea Institute of Science and Technology (KIST), Seoul, Republic of Korea

Abstract

Robots operating in human-centered environments must safely navigate among multiple dynamic obstacles to avoid collisions with people and surrounding infrastructure. Control barrier functions (CBFs) provide an effective mechanism for safety filtering, and recent CBF-based reinforcement learning (RL) methods embed such safety information into learned policies. However, executing only safety-filtered actions during training can restrict policy exploration, a limitation that becomes particularly consequential in dynamic scenes where safety depends on relative robot-obstacle motion. We propose DODGER, a safety-guided RL framework that directly executes policy-generated actions to drive training rollouts while using CBF-filtered references and constraint violations to shape the policy toward collision-avoidance behavior. We evaluate DODGER through a Dubins-car safety analysis and demonstrate goal-directed navigation among multiple dynamic obstacles in full-order humanoid simulation and real-world humanoid experiments using LiDAR-based perception, without a runtime safety filter.

Figures & tables

Fig. 1: A humanoid avoids a moving pedestrian while navigating toward a goal using the DODGER policy. The policy is trained with CBF-based safety guidance and deployed without a runtime safety filter.

Fig. 2: Overview of DODGER. The high-level policy combines local robot-goal information with a graph-based representation of surrounding obstacles to generate navigation commands. During training, the policy action directly drives the environment rollout, while the CBF-filtered reference and constraint violations guide learning. The resulting navigation command is executed by a fixed pretrained locomotion controller.

Fig. 3: Comparison of DODGER and CBF-RL on the Dubins-car benchmark. The dashed black curve denotes the HJ safety boundary

\partial V_{\mathrm{HJ}}

, and the shaded region denotes

\hat{\mathcal{U}}_{\epsilon}

Term	Definition	Weight
$r^{\mathrm{CBF}}$	$\exp(-\lVert\Delta\bm{u}\rVert_{2}^{2}/\sigma^{2})-1$	$100$
$r^{\mathrm{prog}}$	$\Delta\rho$	$60$
$r^{\mathrm{head}}$	$\Delta\bar{e}_{\phi}\exp[-(\rho/\ell_{h})^{2}]$	$10$
$r^{\mathrm{goal}}$	$\mathmybb{1}\{\text{goal reached}\}$	$20$
$r^{\mathrm{obs}}$	$\mathmybb{1}\{\text{obstacle collision}\}$	$-55$
$r^{\mathrm{outside}}$	$\mathmybb{1}\{\bm{p}\notin\mathcal{A}_{\mathrm{hard}}\}$	$-100$

TABLE I: Reward terms used for training DODGER.

Fig. 4: Evaluation success rates of DODGER, CBF-RL-Min, and CBF-RL during the first $300$ million environment transitions for humanoid navigation with obstacle speeds up to $0.6\,\mathrm{m/s}$ . The markers denote individual evaluations, and the solid lines denote the moving average over the three most recent evaluations.

CBF formulation	End (M)	Converged	Stage
DPCBF (DODGER)	113.97	✓	3
C3BF	600	✗	1
Dist.-ECBF	166.20	✓	3
Dist.-ECBF + yaw	154.01	✓	3

TABLE II: Training outcomes under different CBF formulations. “End” denotes the number of environment transitions (millions) at which training terminated; curriculum stages are indexed from 0 to 3.

Metric	DPCBF	Dist.	Dist.+Yaw
Cmd. RMS ( $\mathrm{rad/s}$ )	0.554	0.934	0.894
Cmd. near-sat. (%)	5.48	87.08	61.56
G1 RMS ( $\mathrm{rad/s}$ )	0.435	0.808	0.763
G1 near-sat. (%)	0.87	18.80	15.45

TABLE III: Yaw-rate statistics for different CBF formulations.

Fig. 5: Real-robot navigation with a Unitree G1 among multiple moving pedestrians. The upper panels show the estimated scene and camera views at six representative times, together with DPCBF boundaries and relative velocities; $h<0$ when a relative velocity enters the unsafe side of its corresponding parabolic boundary. The lower-left plot shows the DPCBF values $h$ and margins $d$ of the selected obstacles, with solid curves denoting their minima. Although the minimum $h$ temporarily falls below zero, the minimum $d$ remains positive throughout the experiment.

Component	Configuration
Graph Attention Encoder
Graph input	Robot, goal, and up to $10$ obstacle nodes
Node representation	$\bm{n}_{j}\in\mathbb{R}^{9}$
Edge representation	$\bm{e}_{rj}\in\mathbb{R}^{11}$
Message network $\psi_{1}$	Linear $(11\!\rightarrow\!64)$ + ReLU, Linear $(64\!\rightarrow\!16)$
Attention network $\psi_{2}$	Linear $(16\!\rightarrow\!16)$ + ReLU, Linear $(16\!\rightarrow\!1)$ , masked softmax

TABLE IV: Network architectures used for humanoid navigation.

Explore similar work

Distilling Privileged Control Barrier Functions into RGB-Only Safety Filters for Dynamic Visual Navigation
Sep 29, 2026Seungyeon Yoo, Gawon Lee, Seungwoo Jung +2Obstacle Avoidance Control Barrier Functions
Department of Aerospace Engineering, Seoul National University, Seoul, Republic of Korea
CBF-RL: Safety Filtering Reinforcement Learning in Training with Control Barrier Functions
Oct 16, 2025Lizhi Yang, Blake Werner, Massimiliano de Sa +1Control Barrier Functions Safety Filters
Caltech MCE
Shield-Loco: Shielding Locomotion Policies with Predictive Safety Filtering
Jun 5, 2026Aditya Shirwatkar, Sebastian Sanokowski, Shishir Kolathaya +2Safety Filters Agile Locomotion
Robert Bosch Center for Cyber Physical Systems, Indian Institute of Science, Bangalore, India · Munich Institute of Robotics and Machine Intelligence (MIRMI), Technical University of Munich, Munich, Germany · Department of Computer Science & Automation, Indian Institute of Science, Bangalore, India +2

Sep 29, 2026, cs.ROJ/K move · Enter open · S save

Distilling Privileged Control Barrier Functions into RGB-Only Safety Filters for Dynamic Visual Navigation

Seungyeon Yoo, Gawon Lee, Seungwoo Jung, Inkyu Jang+1

Department of Aerospace Engineering, Seoul National University, Seoul, Republic of Korea

Obstacle Avoidance Control Barrier Functions Safety Filters Object Goal Navigation Teacher-Student Distillation

Open paper AskPDF arXiv

RGB-only end-to-end visual navigation policies remain vulnerable to collisions in real-world dynamic environments, motivating a dedicated safety layer. Existing visual Control Barrier Function (CBF) approaches seek to provide safety from RGB observations, but often rely on real-time rendering or explicit scene reconstruction and are primarily designed for static scenes, limiting their practicality for onboard deployment. We propose a teacher-student visual distillation framework that transfers the safety behavior of a privileged CBF teacher to an RGB-only student filter for dynamic environments. The student maps a short RGB history, robot velocity, and a nominal control action directly to a safe action, while the teacher uses ground-truth robot and obstacle states in a real-to-sim dynamic Gaussian Splatting environment. To reduce the teacher-student information gap, the teacher constructs safety constraints only from obstacles observable within the student's RGB history. It also accounts for obstacle-velocity uncertainty to improve robustness to motion variations, while action augmentation exposes the student to diverse safe and unsafe nominal actions to better capture the safety boundary. At deployment, the student requires only RGB observations and robot velocity, without explicit 3D reconstruction or online rendering. Experiments show that the proposed method outperforms visual CBF baselines and improves the safety of RGB-based navigation policies under dynamic obstacle motion. Project page: https://syeon-yoo.github.io/distill-cbf-site/.

Abstract

Explore similar work

Distilling Privileged Control Barrier Functions into RGB-Only Safety Filters for Dynamic Visual Navigation

CBF-RL: Safety Filtering Reinforcement Learning in Training with Control Barrier Functions

Shield-Loco: Shielding Locomotion Policies with Predictive Safety Filtering