CARVIEW

MOTORHOMES

Select Language

HTTP/2 200 server: GitHub.com content-type: text/html; charset=utf-8 last-modified: Tue, 12 Aug 2025 04:47:48 GMT access-control-allow-origin: * strict-transport-security: max-age=31556952 etag: W/"689ac774-9f5d" expires: Sun, 28 Dec 2025 15:30:38 GMT cache-control: max-age=600 content-encoding: gzip x-proxy-cache: MISS x-github-request-id: C38E:123DE:7BCB78:8AD059:69514AC6 accept-ranges: bytes age: 0 date: Sun, 28 Dec 2025 15:20:38 GMT via: 1.1 varnish x-served-by: cache-bom-vanm7210029-BOM x-cache: MISS x-cache-hits: 0 x-timer: S1766935238.479459,VS0,VE234 vary: Accept-Encoding x-fastly-request-id: 12b1f2f899d7ad6397b78480a5be6f88edca4e3a content-length: 6807 AimBot: A Simple Auxiliary Visual Cue to Enhance Spatial Awareness of Visuomotor Policies

AimBot: A Simple Auxiliary Visual Cue to Enhance Spatial Awareness of Visuomotor Policies

CoRL 2025

Yinpei Dai^{† *}, Jayjun Lee^{‡ *},

Yichi Zhang^†, Ziqiao Ma^†, Jed Yang^†, Amir Zadeh^◊, Chuan Li^◊,

Nima Fazeli^{† ‡ ★}, Joyce Chai^{† ★}

^*Equal Contribution
^★Equal Advising
^†Department of Computer Science and Engineering, University of Michigan, ^‡Robotics Department, University of Michigan,
^◊Lambda Labs

Paper arXiv Code

TL;DR AimBot is a lightweight visual augmentation technique that provides explicit spatial cues (e.g., shooting lines and scope reticles) to improve any Vision-Language-Action (VLA) models.

Abstract

In this paper, we propose AimBot, a lightweight visual augmentation technique that provides explicit spatial cues to improve visuomotor policy learning in robotic manipulation. AimBot overlays shooting lines and scope reticles onto multi-view RGB images, offering auxiliary visual guidance that encodes the end-effector's state. The overlays are computed from depth images, camera extrinsics, and the current end-effector pose, explicitly conveying spatial relationships between the gripper and objects in the scene. AimBot incurs minimal computational overhead (less than 1 ms) and requires no changes to model architectures, as it simply replaces original RGB images with augmented counterparts. Despite its simplicity, our results show that AimBot consistently improves the performance of various visuomotor policies in both simulation and real-world settings, highlighting the benefits of spatially grounded visual feedback.

Model Architecture

AimBot is a simple visual augmentation technique that provides spatial cues in the image space to improve visuomotor policy learning in robotic manipulation.

Experimental Results

Real World Videos Click to view all rollouts

We finetune various base VLA models (OpenVLA+OFT, π₀-FAST, π₀) to evaluate our method.
We compare AimBot against other visual augmentations, including Traces, RoboPoint, and raw depth images.
We visualize sample rollouts below and all rollouts from the paper for all models including OOD scenes are available here.

Performance comparison on five real-world tasks. Each task is evaluated over ten trials.
Model	Fruits in Box	Tennis Ball in Drawer	Bread in Toaster	Place Coffee Cup	Egg in Carton	Total Success
OpenVLA-OFT	7/10	6/10	4/10	2/10	2/10	21/50
OpenVLA-OFT + AimBot	9/10	7/10	9/10	8/10	3/10	36/50
π₀-FAST	10/10	10/10	9/10	7/10	6/10	42/50
π₀-FAST + AimBot	10/10	10/10	10/10	9/10	8/10	47/50
π₀	7/10	7/10	4/10	5/10	4/10	27/50
π₀ + AimBot	10/10	10/10	7/10	8/10	8/10	43/50
π₀ + Traces	8/10	8/10	5/10	2/10	2/10	25/50
π₀ + RoboPoint	8/10	9/10	4/10	6/10	0/10	27/50
π₀ + Depth Images	7/10	9/10	5/10	7/10	4/10	32/50

Task Goal: Put the bread inside the toaster.

Failure: π₀

Success: π₀ + AimBot

Task: Put the cup on the coffee machine.

Failure: π₀

Success: π₀ + AimBot

Task: Put the tennis ball inside the drawer.

Failure: π₀

Success: π₀ + AimBot

Task: Put the eggs inside the egg carton.

Failure: π₀

Success: π₀ + AimBot

Task: Put all the fruits into the box.

Failure: π₀

Success: π₀ + AimBot

Simulation Videos on LIBERO

Performance comparison on the LIBERO simulation benchmark. Green and Red numbers indicate performance gains and losses, respectively. Each task suite averages over four runs.
Model	LIBERO Spatial	LIBERO Object	LIBERO Goal	LIBERO Long	Average Success Rate
OpenVLA-OFT	96.2	97.3	93.9	87.5	93.8
OpenVLA-OFT + AimBot	95.2 (–1.0)	99.1 (+1.8)	94.2 (+0.3)	91.2 (+3.7)	95.0 (+1.2)
π₀-FAST	96.5	96.8	93.6	81.6	92.1
π₀-FAST + AimBot	96.9 (+0.4)	96.8 (+0.0)	94.0 (+0.4)	87.1 (+5.5)	93.7 (+1.6)
π₀	96.8	98.8	95.8	85.2	94.2
π₀ + AimBot	96.9 (+0.1)	98.4 (–0.4)	97.2 (+1.4)	91.0 (+5.8)	95.9 (+1.7)