Efficient robotic data collection remains a major bottleneck for vision-language-action (VLA) policy learning, especially for tasks requiring accurate end-effector positioning and camera-action alignment. This challenge is particularly severe in electronic-waste disassembly, where robots operate in cluttered assemblies of densely packed, visually similar components. Full digital-twin simulation is difficult because many parts are thin, reflective, or mechanically complex, making accurate object-level meshes and contact models difficult to obtain. In this work, we focus on pre-contact reaching and propose a 3D Gaussian Splatting (3DGS)-based data generation pipeline for wrist-camera VLA adaptation. Our method reconstructs the real workspace with 3DGS and renders diverse wrist-camera observations from robot-aligned poses. These observations are paired with smooth, structured reaching trajectories to provide robot-aligned visual-action supervision for VLA fine-tuning. We further augment the VLA policy with an instruction-conditioned reach verifier that estimates target proximity and provides reliable termination before physical interaction. Experiments in motherboard disassembly demonstrate precise VLA-based reaching using wrist-camera observations alone, without an external side-view camera. With π0.5, 1,500 generated trajectories achieve 80.0% precise-reaching success across three targets, compared with 58.7% using 150 teleoperated demonstrations, while requiring less than 10 minutes of active human capture. Project website: https://001-wang.github.io/PrecisionReach-GS/
Figures & tables
Fig. 1: Overview of the proposed real-to-sim-to-real pipeline. The system first captures and calibrates the real workspace, then reconstructs and aligns a 3DGS scene for robot-feasible wrist-camera data generation, and finally fine-tunes and deploys a VLA policy for closed-loop precise reaching.
Method
3DGS
Physics Sim.
Object Mesh
GPU Simulator Req.
SplatSim [ 39 ]
✓
✓
×
✓
RoboGSim [ 36 ]
✓
✓
✓
✓
Real2Render2Real [ 33 ]
✓
×
✓
✓
PrecisionReach-GS
✓
×
×
×
TABLE I: Comparison with representative 3DGS-based robotic data generation methods.
Fig. 2: Rendering representations reconstructed from the same workspace capture. The 3DGS representation better preserves the appearance needed for wrist-camera policy observations.
Fig. 3: Overview of the instruction-conditioned reach verifier. The verifier integrates the current wrist-camera observation, language instruction, and robot state to estimate whether the end-effector has reached the target interaction region. Language-conditioned FiLM and spatial attention adapt the visual representation to the instructed target, while state masking discourages reliance on robot-state shortcuts. During deployment, reaching is confirmed only after three consecutive valid predictions.
Fig. 4: Conceptual illustration of progressive wrist-camera feedback for precise reaching. The robot first performs coarse target localization, then refines its motion through repeated visual feedback, and finally stops when the reach verifier confirms alignment with the target interaction region.
Fig. 5: Qualitative progression of three real-robot reaching trajectories (top to bottom), each shown as six frames from the initial configuration to the final target interaction pose.
Fig. 6: Real-robot reaching performance on a physically repositioned motherboard workspace that differs from the scene used for 3DGS reconstruction and data generation. Detailed trial counts are reported in Table II .
Policy
Target
Metric
Teleop Trajectories
PrecisionReach-GS Trajectories
50
100
150
300
900
1500
π0.5
Screw
Coarse reach
10/25
20/25
25/25
5/25
23/25
25/25
Precise reach
3/25
10/25
12/25
0/25
14/25
21/25
GPU
Coarse reach
8/25
23/25
25/25
15/25
23/25
25/25
Precise reach
4/25
11/25
14/25
0/25
14/25
19/25
RAM
Coarse reach
10/25
22/25
25/25
16/25
24/25
25/25
TABLE II: Real-robot reaching performance under different amounts of teleoperation and PrecisionReach-GS training data. Each entry reports successful trials out of 25. Coarse reach measures arrival in the target neighborhood; precise reach additionally requires a terminal pose from which the fixed grasp-and-lift routine succeeds.
Method
# Traj.
Real Capture
Active Human Time
Total Time
Teleoperation
50
50 demos
∼ 90 min
∼ 90 min
100
100 demos
∼ 150 min
∼ 150 min
150
150 demos
∼ 200 min
∼ 200 min
PrecisionReach-GS
300
4 videos
<5 min
∼ 45 min
900
12 videos
<10 min
∼ 120 min
1500
12 videos
<10 min
∼ 200 min
TABLE III: Data acquisition cost and trajectory smoothness for teleoperation and PrecisionReach-GS. For PrecisionReach-GS, total time is measured for a single generation window. RMS jerk is reported in m/s3 ; lower is better. Detailed jerk statistics are reported in Table V .
Target
Initial Visual Context
300 Traj.
1500 Traj.
Coarse SR
Precise SR
Coarse SR
Precise SR
Screw
Full workspace context
4/10
0/10
10/10
10/10
Limited workspace context
1/10
0/10
9/10
9/10
No workspace context
0/10
0/10
6/10
2/10
GPU
Full workspace context
8/10
0/10
10/10
9/10
Limited workspace context
6/10
0/10
9/10
9/10
TABLE IV: Initial-view robustness under different generated data scales and target components. Starting views are grouped by the amount of workspace context visible in the wrist-camera observation.
Method
Mean ↓
Median ↓
P95 ↓
Max ↓
Teleoperation
3.2657
3.2532
4.1228
4.5666
PrecisionReach-GS
0.2353
0.1683
0.6969
0.9853
TABLE V: Trajectory smoothness comparison between teleoperation and PrecisionReach-GS. RMS jerk is reported in m/s3 ; lower values indicate smoother trajectories.
Fig. 7: Representative recovery trajectory under the no-workspace-context condition. The top row shows the robot pose and the bottom row shows the corresponding wrist-camera observation. Initially, the motherboard is completely outside the wrist camera’s field of view. As the robot moves, the workspace gradually becomes visible, allowing the policy to continue toward the target region.
Controlled factor
Original setting
Matched setting
Precise success (%)
Observation modality
Teleop-150, wrist+side
Teleop-150, wrist only
58.7→53.3
Trajectory timing
GS-900, smoothstep
GS-900, linear
58.7→50.7
TABLE VI: Controlled analysis with π0.5 (25 trials per target).
Metric
Full
No verifier
No KF weighting
Coarse success (%)
93.3
93.3
74.7
Precise success (%)
58.7
45.3
46.7
TABLE VII: Ablation of key pipeline components with 900 generated episodes.
Verifier
MAE (cm) ↓
Near-Target MAE (cm) ↓
Stop Prec. (%) ↑
Biased-State Prec. (%) ↑
Full
0.64±0.02
0.10±0.02
99.8
48.1±11.6
w/o spatial attention
0.64±0.07
0.12±0.02
99.8
–
w/o FiLM
0.67±0.06
0.12±0.03
99.5
–
w/o instruction conditioning
0.71±0.05
0.12±0.03
99.3
22.8±16.1
w/o state masking
0.44±0.01
0.08±0.01
100.0
0.0
TABLE VIII: Ablation and robustness analysis of the reach verifier. Distance metrics are evaluated using the correct target instruction. Near-target MAE is computed over frames whose ground-truth target distance is below 3 cm. State robustness is evaluated by applying a −2 cm bias to the end-effector height.
Fig. 8: Comparison between real and rendered observations.
Metric
Mean
Best
Worst
SSIM
0.682
0.919
0.503
Correlation
0.843
0.964
0.750
MSE similarity
0.979
0.991
0.963
TABLE IX: Visual similarity between real and rendered robot videos.
Metric
S1
S2
S3
Position RMSE (m)
0.0114
0.0094
0.0096
Position mean error (m)
0.0102
0.0077
0.0076
Position max error (m)
0.0258
0.0326
0.0329
Anchor position mean error (m)
0.0039
0.0032
0.0040
Anchor position max error (m)
0.0047
0.0042
0.0067
Orientation mean error (deg)
0.5792
0.4894
0.4614
TABLE X: Robot-frame alignment accuracy after transforming the reconstructed scene into the robot coordinate frame.
Despite recent advances in Vision-Language-Action models (VLAs) for robotic manipulation, their performance remains sensitive to changes in camera configuration. The problem becomes more evident in cross-setup deployment, as reproducing the exact camera pose used for training is nearly impossible. Unlike fixed external views, wrist views are more challenging because the camera moves with the robot, causing even small mounting variations to alter fine-grained geometric cues. To address this, we propose WARP-VLA, a camera-view robust VLA for diverse wrist camera configurations. WARP-VLA adopts a Mixture-of-Experts (MoE) architecture where individual experts learn view-specific feature transformations, and a router combines them based on implicit view information. This allows the policy to be deployed without requiring camera extrinsic parameters as additional input. Through experiments on the LIBERO benchmark, WARP-VLA improves the average success rate of pi-0.5 from 39.2% to 78.3% under wrist-view perturbations. The real-robot experiments further show that the feature-level adaptation learned in simulation successfully transfers to diverse deployment settings. To facilitate reproducibility and future research, we release our wrist viewpoint robustness benchmark and a plug-and-play implementation.
Junmyeong Lee, Dongmin Shin, Min-Gyu Park +3
Yonsei University. · Korea Electronics Technology Institute(KETI).
Vision-Language-Action (VLA) models have emerged as powerful foundations for robotic manipulation, but their reliance on fixed camera configurations during training makes them brittle to changes in camera count or pose during deployment. To overcome these limitations, we propose VersaCamVLA, a camera-configurable framework that decouples camera-set representation from action learning. VersaCamVLA learns a unified scene-token interface that maps an arbitrary, variable set of posed RGB views into fixed-size latent scene tokens. This is achieved via multi-signal target-view prediction and Wrist-Augmented Pose Sampling (WAPS), which leverages natural wrist-camera motion for free pose diversity. At deployment, a lightweight spatial encoder injects these compact scene tokens into a pretrained base VLA as a supplementary visual condition, requiring no explicit 3D sensing or novel-view rendering. Experiments on RoboTwin, LIBERO, and a real-robot platform demonstrate that VersaCamVLA consistently outperforms prior VLA methods and direct multi-view baselines, maintaining robust performance across varying camera counts and unseen camera poses.
Boyao Han, Chen Shi, Jingjing Qian +2
The Chinese University of Hong Kong, Shenzhen · Shenzhen Loop Area Institute · Harbin Institute of Technology, Shenzhen
Current Vision-Language-Action (VLA) models often struggle with high-precision robotic manipulation. We attribute this limitation primarily to their visual attention being dispersed across task-irrelevant regions. To address this issue, we propose ActGaze, a training approach that guides VLA policies to gaze on task-relevant regions, much like humans gaze on critical visual cues while executing precise movements. Unlike prior methods that rely on external labels for gaze supervision, ActGaze derives spatial supervision directly from the VLA's own action objective by using counterfactual visual interventions to identify regions that are critical for action prediction. Extensive real-robot experiments on four high-precision robotic manipulation tasks demonstrate that ActGaze induces more focused visual attention on task-relevant regions and consistently outperforms the base VLA policy and other visual-grounding approaches.
Jinxuan Zhu, Jiaheng Wang, Chao Tang +7
Samsung Robotics eXperience · National University of Singapore