cs.ROMar 12, 2026
SaveLearning Visuomotor Policy for Multi-Robot Laser Tag Game
Organizations: College of Computer Science and Technology at Zhejiang University, Hangzhou, China. · School of Engineering at Westlake University, Hangzhou, China.
Abstract
In this paper, we study multi robot laser tag, a simplified yet practical shooting-game-style task. Classic modular approaches on these tasks face challenges such as limited observability and reliance on depth mapping and inter robot communication. To overcome these issues, we present an end-to-end visuomotor policy that maps images directly to robot actions. We train a high performing teacher policy with multi agent reinforcement learning and distill its knowledge into a vision-based student policy. Technical designs, including a permutation-invariant feature extractor and depth heatmap input, improve performance over standard architectures. Our policy outperforms classic methods by 16.7% in hitting accuracy and 6% in collision avoidance, and is successfully deployed on real robots. Code will be released publicly.
Figures & tables
Fig. 1: Illustration of the multi-robot laser tag game. The laser beams shown are virtual for visualization purposes.
Fig. 2: This figure shows the pipeline of our method. The teacher policy takes privileged states and outputs velocity command, and is trained with MARL. The student policy takes a time series of images and outputs velocity command. The student policy imitates the action output of the teacher and is deployed onboard. Shape of each tensor in the policy is shown. denotes the number of historical images used for the recurrent module. DATv2 is the monocular depth estimation method Depth Anything v2 [ 19 ] .
Fig. 3: Structure of the feature extractor of the teacher policy. Opponent, obstacle and teammate states are encoded with self-attention and summation pooling. Each entity token (White circle) in the figure represents one embedded instance of the obstacle or neighbor robot.
Fig. 4: The original image is sent into YOLOv5 and DATv2 to generate detection results and depth images. The detection results contain two classes, namely the enemy and the ally. A heat map that indicates the detection location is generated for each class. Then the two classes of heat maps are concatenated with the depth image along the channel axis to form the input tensor.
| Visual Encoders | Input Types | Simulation Train | Simulation Test | Real-World Test | Time (ms) | ||||||
| AE | Hit@15 | CR[%] | AE | Hit@15 | CR [%] | AE | Hit@15 | CR [%] | |||
| ResNet50 [ 23 ] | RGB | 0.0012 | 25.8 ± 6.4 | 40.0 | 0.0075 | 14.3 ± 4.0 | 66.0 | 0.0115 | 0.0 ± 0.0 | 70.0 | 21.6 (28.1) |
| CLIP RNet [ 24 ] | RGB | 0.0013 | 21.8 ± 3.0 | 54.0 | 0.0066 | 6.4 ± 2.4 | 58.0 | 0.0102 | 0.0 ± 0.0 | 54.0 | 20.1 (25.4) |
| DINOv2 [ 25 ] | RGB | 0.0020 | 17.4 ± 3.0 | 30.0 | 0.0078 | 6.1 ± 2.6 | 40.0 | 0.0228 | 0.0 ± 0.0 | 50.0 | 137.2 (185.7) |
| MAE [ 26 ] | RGB | 0.0009 | 33.8 ± 1.8 | 38.0 | 0.0015 | 19.7 ± 3.0 | 42.0 | 0.0099 | 7.8 ± 5.2 | 44.0 | 24.9 (28.0) |
| SimpleCNN | Edges [ 11 ] | 0.0006 | 26.8 ± 2.8 | 38.0 | 0.0020 | 20.8 ± 4.1 | 46.0 | 0.0032 | 12.5 ± 5.9 | 72.0 | 7.8 (18.2) |
TABLE I: Comparison and ablation of visual encoders and input types, evaluated by Action Error (AE), Collision Rate (CR), and Hit@15. SimpleCNN refers to a lightweight CNN with channel sizes [16, 32, 32] and filter sizes [5, 4, 3]. Onboard inference time (mean with 99% percentile in parentheses) is reported only for the visual encoder. “train” means in the same training environment, and “test” means in unseen test environments. We compare RGB, edge images [ 11 ] , depth, and our input (depth + heatmaps, see Fig. 4 ).
Fig. 5: The first row shows enemy and ally detection results, the second row shows depth maps, and the third row shows full-gradient attention maps [ 29 ] . The attention maps highlight task-relevant regions of the input image, typically focusing on enemy targets and obstacles, consistent with human intuition. The first three columns present simulation data, and the last column presents real-world data.
| Action Methods | AE | Hit@15 | Hit@30 | CR [%] |
| Teacher-MLP | N/A | 40.6 ± 1.6 | 19.0 ± 1.7 | 6.0 |
| Teacher-SA+SDF [ 7 ] | N/A | 40.9 ± 2.4 | 17.6 ± 2.0 | 6.0 |
| Teacher-SA [ 6 ] | N/A | 42.3 ± 2.0 | 18.0 ± 2.9 | 10.0 |
| Teacher-SAP (Ours) | N/A | 48.6 ± 1.7 | 22.1 ± 2.0 | 10.0 |
| BC-MLP | 0.0251 | 0.0 ± 0.0 | 0.0 ± 0.0 | 80.0 |
| BC-LSTM | 0.0059 | 5.3 ± 1.7 | 0.0 ± 0.0 | 48.0 |
TABLE II: The gray rows are results of different teacher policies, and the white are those of the students. Teacher-SAP refers to the teacher policy with our self-attention-and-pooling as the feature extractor. Teacher-SA refers to policy with vanilla self-attention feature extractor used in MARL [ 6 ] without permutation-invariant operators. Teacher-SA+SDF refers to [ 7 ] ’s feature extractor design. BC denotes behavioral cloning.
Fig. 6: The first row shows the trajectories of robots in the laser tag game in simulation. When the firing condition is met, an ally robot fires a laser beam toward the enemy, indicated by the red line. The second row shows the robot motions, with enemies in red and allies in white. Allies maneuver strategically to aim and attack the enemies, while the enemy maneuver to escape.
Fig. 7: Snapshots from the real-world experiment.
Fig. 8: Left : Cumulative distribution function (CDF) of radial distances from the crosshair center. A steeper rise near the origin indicates that more shots are concentrated close to the crosshair, reflecting better laser aiming accuracy. Right : Hit@15 score under different maximum velocities. Both figures illustrate that ours have an improved laser hitting performance.
| Methods | Hit@15 | Hit@30 | CR[%] | |
| State Estimation | Control Policy | |||
| GT | PD | 44.0 ± 1.8 | 20.0 ± 1.7 | 34.0 |
| GT | MLP [ 16 ] | 44.2 ± 3.0 | 20.2 ± 1.3 | 50.0 |
| GT | IBVS [ 30 ] | 35.2 ± 5.5 | 16.0 ± 5.1 | 68.0 |
| GT | Ours-Teacher | 48.6 ± 1.7 | 22.1 ± 2.0 | 10.0 |
| BA [ 4 ] | PD | 33.0 ± 1.8 | 13.9 ± 2.4 | 42.0 |
TABLE III: Comparison with other methods. GT uses ground-truth states as input. Ours-Teacher refers to the state-based teacher policy, and Ours-Student to the vision-based student policy. Gray rows indicate methods relying on onboard sensing.
Fig. 9: Relative distance and heading between the ally and enemy in one episode, with red segments marking when the enemy is within the shooting zone. The visuomotor policy yields more successful targeting frames than the PD policy.
More in this topic
Visuomotor Policy Learning
Visuomotor policy learning trains robots to turn camera observations, often with language and proprioception, into actions for manipulation and navigation.
See all 286 papers