LEGS: Fine-Tuning Teleop-Free VLAs for Humanoid Loco-manipulation in an Embodied Gaussian Splatting World
LEGS: Fine-Tuning Teleop-Free VLAs for Humanoid Loco-manipulation in an Embodied Gaussian Splatting World
This research presents LEGS, a hybrid simulator that combines a mesh foreground with a photorealistic 3D Gaussian Splatting background to train vision-language-action (VLA) policies for humanoid loco-manipulation without human teleoperation. The method uses a procedural motion-primitive generator to synthesize labeled demonstrations and a two-stage color calibration to align renders with the deployment camera. Policies trained on LEGS data match or exceed those trained on human teleoperation demos and outperform mesh-only baselines. The approach enables cost-effective re-rendering of demonstrations for new scenes and objects.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Unitree G1 humanoid robot
1
View original evidence
-and-place tasks of increasing whole-body difficulty and three VLA backbones ($\psi_0$, $\pi_{0.5}$, GR00T N1.6), a policy trained purely on LEGS data matches or exceeds one trained on human teleoperation demos on every experiment. It also outperforms a mesh-only baseline trained on the same motions without the calibrated 3DGS frontend. Because motion is recorded independently of appearance, the same demonstrations c
Open source S4pick-and-place tasks
3
View original evidence
-and-place tasks of increasing whole-body difficulty and three VLA backbones ($\psi_0$, $\pi_{0.5}$, GR00T N1.6), a policy trained purely on LEGS data matches or exceeds one trained on human teleoperation demos on every experiment. It also outperforms a mesh-only baseline trained on the same motions without the calibrated 3DGS frontend. Because motion is recorded independently of appearance, the same demonstrations c
Open source S4VLAs backbones
3
View original evidence
-and-place tasks of increasing whole-body difficulty and three VLA backbones ($\psi_0$, $\pi_{0.5}$, GR00T N1.6), a policy trained purely on LEGS data matches or exceeds one trained on human teleoperation demos on every experiment. It also outperforms a mesh-only baseline trained on the same motions without the calibrated 3DGS frontend. Because motion is recorded independently of appearance, the same demonstrations c
Open source S4policy trained on LEGS data
1
View original evidence
-and-place tasks of increasing whole-body difficulty and three VLA backbones ($\psi_0$, $\pi_{0.5}$, GR00T N1.6), a policy trained purely on LEGS data matches or exceeds one trained on human teleoperation demos on every experiment. It also outperforms a mesh-only baseline trained on the same motions without the calibrated 3DGS frontend. Because motion is recorded independently of appearance, the same demonstrations c
Open source S4
- code: Available Artifact link Open source S5
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
arXiv:2606.01458v2 Announce Type: replace Abstract: Training vision-language-action (VLA) policies for humanoid loco-manipulation is constrained by the high cost and complexity of collecting human teleoperation demonstrations. VLA policies fine-tuned in simulators have so far transferred poorly to humanoid loco-manipulation tasks. We present LEGS (Loco-manipulation via Embodied Gaussian Splatting), a hybrid simulator
Open source S2
that composites a mesh foreground (robot, objects, props) over a photorealistic 3D Gaussian Splatting (3DGS) background reconstructed from a handheld scene capture. LEGS uses a procedural motion-primitive generator to synthesize labeled demonstrations at scale without human teleoperation, and a two-stage color calibration that aligns renders to the deployment camera. On a Unitree G1 humanoid robot, across three pick
Open source S3
-and-place tasks of increasing whole-body difficulty and three VLA backbones ($\psi_0$, $\pi_{0.5}$, GR00T N1.6), a policy trained purely on LEGS data matches or exceeds one trained on human teleoperation demos on every experiment. It also outperforms a mesh-only baseline trained on the same motions without the calibrated 3DGS frontend. Because motion is recorded independently of appearance, the same demonstrations c
Open source S4
an be re-rendered for a target scene and objects at over 15x lower cost than teleoperation. When both the objects and the scene change, a policy trained on LEGS-AUG data re-rendered for that target maintains task success, while the teleop baseline fails entirely. Code and videos are available at https://hojjunekim.github.io/legs/.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org