VLAct (StarVLA)

VLAct is a continued pre-training method for vision-language-action robot policies, building a reusable Qwen3-VL-4B action backbone from open data.

Made by StarVLA.

Licenses

Code: MIT (opens in a new tab)
Read by Robotheca from LICENSE (opens in a new tab) (MIT) at commit 4c41b66 (opens in a new tab), on 16 Oct 2026.
Weights: Apache-2.0 (opens in a new tab)
Read on huggingface.co/StarVLA/VLAct_Qwen3_Pretrain (opens in a new tab).
Commercial use: Yes

Freezes the vision encoder and lower LLM layers and mixes image-caption data into training to preserve the VLM's broad visual knowledge.
Trains three continuous action heads (OFT, PI, GR00T) on one shared latent so the backbone cannot rely on a single decoder's features.
Uses a partially unified 20-dimensional action layout across embodiments, with a wrap-aware loss for periodic joint angles.
Reports results on LIBERO-Plus, VLA-Arena, RoboTwin 2.0, DOMINO, RoboCasa-GR1 and the RoboDojo leaderboard.
Provides launch scripts for single-node and Slurm multi-node pre-training, downstream fine-tuning and benchmark evaluation.
Publishes the pre-trained backbone and fine-tuned checkpoints for several benchmarks on Hugging Face.

Written by Robotheca from the project's README (opens in a new tab) at commit 4c41b66.

Difficulty: Advanced
Hardware it needs, as the project states it: Tested workflows assume Linux, Python 3.10, NVIDIA GPUs and a CUDA-compatible PyTorch build. The pre-training launcher uses 8 GPUs on one node or multiple nodes via Slurm; the Franka fine-tuning used 8 H800 GPUs.
How to start: The project's website (opens in a new tab)

Checked on GitHub on 16 Oct 2026.

No robot on Robotheca is linked to it yet. Links come from a robot's robot.yaml, from its dependency files read as text, or from Robotheca's editors.

Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.