Autonomous mobile robots require spatial representations that capture geometry and semantics while staying sparse enough for planning in large, unknown environments. Graphs provide a natural representation for this purpose and underpin both classical planning methods and a growing class of reinforcement-learning (RL) policies. However, RL training needs hundreds to thousands of parallel environments, each constructing and updating its own graph. Existing graph libraries are typically designed for either single-robot deployment or fixed graphs, and CPU-based graph construction limits scalable training to only a few dozen environments. We introduce GGraph, an open-source GPU-accelerated graph library for both robot deployment and large-scale learning. GGraph supports real robots through ROS 2 interfaces and vectorized GPU-based policy training, with built-in support for tasks including 3-D navigation, exploration, and loop-closure integration. For single-robot operation, GGraph updates graphs in milliseconds. At scale, it enables graph-based exploration policies to train across hundreds of parallel robots, reducing wall-clock training time by one order of magnitude. Across simulation and real-world experiments, these results establish GGraph as a unified graph infrastructure for both large-scale robot learning and deployment.
Figures & tables
Fig. 1 : GGraph is a GPU-accelerated graph library for robot deployment and learning. The image shows 200 robots training graph-based exploration policies in parallel in Isaac Sim. Nodes (red) are sampled in 3-D free space and edges (colored per robot) show collision-free connectivity.
Library
GPU Accel.
Spatial Graph Merging
Robotics Navigation & Exploration
Policy Learning Support
NetworkX
✗
✗
✗
✗
cuGraph [ 13 ]
✓
✗
✗
✗
PyTorch Geometric [ 12 ]
✓
✗
✗
✓
GBPlanner [ 7 ]
✗
✓
✓
✗
TARE Planner [ 6 ]
✗
✓
✓
✗
ARiADNE node manager [ 9 ]
✗
✓
✓
✓
TABLE I : Comparison of graph libraries and planners across the capabilities a navigation graph needs for deployment and learning.
Fig. 2 : GGraph constructs 3-D navigation graphs from real-world data captured with an Odin camera. Local point clouds are converted into occupancy grids, from which the library samples nodes (red). The inset shows the graph projected into the camera view. The traversable structure of a multi-level outdoor scene is stored in a few thousand nodes, whereas a volumetric map of the same scene would require orders of magnitude more memory.
Fig. 3 : Node attributes for exploration and navigation. Bird’s-eye view of one region of the campus graph of Fig. 2 . (A) The frontier layer marks nodes on the boundary between explored and unexplored space (yellow) against free-space nodes (red). (B) A semantic layer marks nodes on grass or bushes (green) against paved ground (red), using RGB input.
Fig. 4 : Loop closure on a GGraph graph. The graph of a drifted square trajectory (A) before and (B) after the correction. Dashed: true path; solid: estimated path.
Fig. 5 : Graph update time vs number of nodes. Each point is one GGraph update. A local graph built from an 80 × 80 m observed window is merged into a global graph of N nodes, with collision checks on the new edges. Lines show the total and per-stage update time, averaged over three obstacle layouts.
Planner
Graph builder
Easy
Medium
Complex
All
Success rate
Nearest frontier (greedy)
ARiADNE
1404
1896
2388
1896
0.44
ARiADNE SAC-16 (6 h)
ARiADNE
909
1304
2305
1506
0.86
ARiADNE released policy (24 h, 32 robots)
ARiADNE
861
1090
2050
1333
0.91
SAC-512 (6 h)
GGraph
736
1205
2165
1369
0.98
PPO ( 512×32 , 6 h)
GGraph
785
1156
1825
1255
0.97
TABLE II : Performance of the learned exploration policy on 150 validation maps (50 per difficulty tier). Values are mean path length in grid cells (px). Success is the fraction of maps in which complete coverage was achieved.
Fig. 6 : Throughput and rollout allocation. (A) Complete environment-update throughput for GGraph and ARiADNE as the number of parallel robots increases; both axes are logarithmic. (B) PPO reward versus training wall-clock time. Each configuration collects NT=16,384 transitions per iteration, with N environments and T steps per environment.
Fig. 7 : Training curves showing exploration performance, measured by the path length required to complete exploration, versus wall-clock training time. Lower path length is better. The dashed line marks the performance of ARiADNE’s released ckpt.
Fig. 8 : Exploration trajectories on medium (top) and complex (bottom) validation maps. PPO takes the shortest path among the learned policies in both examples, while the nearest-frontier baseline fails to complete either map. White circles mark starts, red squares mark final poses, and gray regions remain unknown. Labels report path length; DNF denotes failure to complete exploration.
Fig. 9 : Exploration of the forest world (150 × 150 m) by a ground vehicle in the CMU autonomous exploration development environment [ 30 ] . Left to right: ARiADNE’s recipe trained for 2 h with 16 Ray workers; TARE [ 6 ] ; ARiADNE’s released checkpoint trained for 24 h; our SAC policy trained for 2 h with 512 robots on GGraph ; and our PPO policy trained for 1.5 h with 512 robots on GGraph . Each panel is a run from the same start pose, and the numbers are the distance driven and the time taken for completion. White circle: start; red square: end. As another baseline, GRATE [ 11 ] reports a best path length of 1026 m under its own completion rule, using a policy trained for about 36 h. Our PPO approach on GGraph achieves similar performance in just 1.5 h of training.
Fig. 10 : Exploration of the large-scale indoor environment [ 30 ] by our PPO-based policy trained using GGraph for 1.5 h. It achieves a path length of 1048±64 m across 10 seeds. The reported baseline for exploration in the same map by ARiADNE is 1042 m using a policy trained with SAC for 24 h.
Fig. 11 : The simulation-trained PPO policy successfully explores a real indoor map on the Spot robot using on-board cameras only. The top-down point cloud (colour = height) with the 157m trajectory is shown. Insets: an image of the corridor and the robot.
Many robotic exploration algorithms rely on graph structures for frontier-based exploration and dynamic path planning. However, these graphs grow rapidly, accumulating redundant information and impacting performance. We present a hybrid transformer-based framework trained with Proximal Policy Optimization which complements exploration algorithms by pruning these graphs during exploration, limiting their growth and reducing the accumulation of excess information. The framework was evaluated on simulations of a robotic agent using Rapidly-Exploring Random Trees to carry out frontier-based exploration, where the learned policy reduces graph size by up to 96%. We find preliminary evidence that our framework enables effective, generalizable exploration under reduced information density, consistently outperforming the randomly pruned baseline and improving both exploration efficiency and computational efficiency beyond the unpruned baseline for complex long-horizon exploration tasks. To the best of our knowledge, these results are the first suggesting the viability of RL to prune dynamic graphs used in robotic exploration algorithms.
Adithya V. Sastry, Bibek Poudel, Weizi Li
Independent Researcher Knoxville, TN, USA · Purdue University West Lafayette, IN, USA · University of California, Riverside Riverside, CA, USA
For robots to work reliably in commercial and industrial applications, recent advances in agentic coding systems have potential to combine interpretable robot programming with the open-world adaptability of model-free policies. We focus on "Variational Automation" (VA), a class of tasks that have larger variations in object geometry and pose than fixed automation. Model-free policies often struggle to close the reliability gap for VA tasks, which must be executed persistently and reliably in commercial and industrial applications. Motivated by prior work on Task and Motion Planning (TAMP) and the Robot Operating System (ROS), we introduce Graph-as-Policy (GaP), a multi-agent coding harness that generates directed computation graphs with perception, planning, and control nodes from a Modular Open Robot Skill Library (MORSL). GaP generates an internal simulation environment to rehearse task instances with different graphs in parallel to iteratively refine the graph structure and parameters to improve success rates and throughput. Evaluation with 8 new open VA task benchmarks, 4 in simulation and 4 in real-world, suggests that GaP can achieve success rates that significantly outperform baselines. Details, code, and data: https://graph-robots.github.io/gap
Kaiyuan Chen, Shuangyu Xie, Letian Fu +21
University of California, Berkeley · Carnegie Mellon University · Bosch +1
Training Deep Reinforcement Learning (DRL) navigation policies for different robot configurations remains time-consuming. We present FlashNav, a GPU-based framework that trains robot-specific navigation policies within tens of seconds. A unified robot specification configures a lightweight simulator for batched motion updates, range sensing, and footprint collision checking over a shared occupancy map. The framework supports nonconvex footprints, different sensor configurations and drive types. Blockwise ray queries and selective observation recomputation after resets reduce simulation overhead, while GPU-resident replay and overlapping experience collection and learner updates support efficient off-policy training. Experiments covered five robot configurations and three computing platforms. With FastDSAC on a single RTX 5090 GPU, FlashNav can train a deployable navigation policy in under 30 seconds. FlashNav achieved the highest success rate and score in the benchmark comparison. The selected policies were deployed on wheeled, quadrupedal, humanoid, and irregularly shaped robots without additional policy training.
Shanze Wang, Yiwei Qian, Xinming Zhang +8
Eastern Institute of Technology, Ningbo · The Hong Kong Polytechnic University · National University of Singapore +2