Robot learning / 3D Gaussian Splatting

PrecisionReach-GS

Robot-Aligned 3D Gaussian Splatting for Precise Reaching with Wrist-Camera VLA Policies

A lightweight real-to-sim-to-real pipeline that turns one reconstructed workspace into scalable, robot-aligned wrist-camera demonstrations for precise pre-contact reaching.

Zuoxu Wang, Xiao Liang^* Texas A&M University * Corresponding author

Read the paper Project repository

Real-robot deployment Wrist-camera policy · 16× playback

1,500generated trajectories

<10 minactive human capture

0.235generated RMS jerk

80%π_0.5 precise reach at 1,500 trajectories

01The problem

Precise reaching needs better data, not a heavier simulator.

Dense electronic assemblies leave little room for camera-action mismatch. Teleoperation is slow and noisy, while full digital twins demand meshes, collision geometry, contact models, and extensive setup.

Research question

Can a lightweight 3DGS reconstruction train a wrist-camera VLA policy to reach small targets precisely?

Abstract

Efficient robotic data collection remains a major bottleneck for vision-language-action policy learning, especially when tasks require accurate end-effector positioning and reliable camera-action alignment. PrecisionReach-GS focuses on the pre-contact reaching stage and reconstructs the real workspace with 3D Gaussian Splatting. It renders diverse wrist-camera observations from robot-aligned poses, pairs them with smooth reaching trajectories, and uses the resulting data to fine-tune VLA policies.

Experiments in motherboard disassembly show precise wrist-camera-only reaching without requiring an external side-view camera in this task setting. The pipeline reduces active human collection effort, produces smoother supervision, and avoids building a complete physical digital twin.

02Method

From a real workspace to robot-aligned supervision.

The method reconstructs the scene once, aligns it with the robot frame, then renders visual-action pairs along feasible reaching trajectories.

PrecisionReach-GS pipeline from real-world capture through robot-aligned 3DGS data generation to VLA fine-tuning and real-robot deployment

Open full-resolution workflow

Capture and reconstruct

Short real-world captures preserve the workspace appearance, clutter, and target layout in a photorealistic 3DGS scene.

Align and generate

A weighted similarity transform maps the reconstruction into the robot frame. Feasible start poses and smooth trajectories produce synchronized wrist views, states, and actions.

Fine-tune and reach

The VLA policy predicts closed-loop reach actions from wrist images and language instructions. A lightweight verifier decides when the target region has been reached.

03Data engine

The reconstructed workspace becomes a camera-action generator.

Camera motion stays tied to feasible end-effector motion, so every rendered wrist observation receives a matching robot-frame action label.

Robot-aligned generationPose sampling, rendering, and trajectory export in one workflow

Scene-scale viewPhotorealistic workspace reconstruction

Policy viewRendered observations from robot-feasible wrist poses

04Quantitative results

More generated data produces more precise real-robot reaches.

Tests use a physically repositioned motherboard, so the policy must tolerate moderate target-location changes rather than replay a fixed pose.

TeleoperationPrecisionReach-GS

Each value averages screw, GPU, and RAM results. Every target uses 25 real-robot trials.

21 / 25

Screw precise reach

π_0.5 trained with 1,500 generated trajectories.

19 / 25

GPU precise reach

The target requires more wrist rotation near the interaction pose.

20 / 25

RAM precise reach

Wrist-camera-only deployment in a repositioned workspace.

Real-robot coarse and precise reaching performance for Pi 0.5 and OpenVLA-OFT across teleoperation and PrecisionReach-GS data scales

Real-robot performance by data scale

29 / 30precise reaches from full-context initial views

27 / 30precise reaches from limited-context initial views

17 / 30coarse recovery when the board starts out of view

58.7%precise success with the full 900-trajectory pipeline

05Real-robot demos

Precise reaching after the motherboard moves.

Select a target to watch the wrist-camera VLA policy close the loop on a physical Franka robot. Videos play at 16× speed.

01GPU precise reach

Closed-loop reaching under a changed target-board location.

Low-context start

The first wrist image can contain only the table surface.

With 1,500 generated trajectories, the policy recovers the target neighborhood in 17 of 30 trials even when the motherboard begins outside the wrist camera’s field of view.

Robot viewExternal camera · 6×

Wrist-camera viewPolicy observation

06Why 3DGS

Visual realism without an object-level digital twin.

Targets, clutter, and workspace appearance remain visible in the Gaussian representation, while conventional mesh reconstruction struggles with thin and reflective components.

3D Gaussian Splatting rendering of the motherboard workspace — 3DGSDense visual context

Mesh reconstruction of the same motherboard workspace with missing geometry — MeshIncomplete thin geometry

07Resources

Build on PrecisionReach-GS.

The repository includes the robot-frame alignment, dataset preparation, trajectory conversion, and browser-based Gaussian Splat data-generation tools used by the pipeline. Large datasets and trained models remain external.

01 / ALIGN

Robot-frame calibration

Fit a weighted COLMAP-to-robot similarity transform and export aligned camera poses for the viewer.

02 / PREPARE

3DGS dataset tools

Prepare robot-pose-aware inputs and convert recorded camera paths into end-effector supervision.

03 / GENERATE

Dual-view GS viewer

Inspect aligned splats, sample reachable starts, build paths, and record synchronized visual-action data.

Paper PDF GitHub repository

@unpublished{wang_precisionreach_gs,
  title  = {Robot-Aligned 3D Gaussian Splatting for Precise Reaching
            with Wrist-Camera VLA Policies},
  author = {Wang, Zuoxu and Liang, Xiao},
  note   = {Manuscript under review},
  url    = {https://github.com/001-Wang/PrecisionReach-GS}
}