Unitree's G1 shifts to useful manipulation
Unitree’s G1 robot has moved beyond flashy demos toward practical object manipulation and tool use by leveraging an open-source model (UnifoLM-VLA-0) to extend dexterity and generalization reported. The trend toward open, general-purpose manipulation models mirrors how foundation models spread across perception and control.
Unitree published the training and inference code plus model weights for UnifoLM‑VLA‑0 on Jan 29, 2026. (github.com) UnifoLM‑VLA‑0 is presented as a Vision–Language–Action (VLA) large model trained end‑to‑end on a real‑robot dataset covering 12 manipulation categories and able to run a single multi‑task policy across them. (github.com) Unitree’s released experiments and dataset names show concrete tasks such as opening/closing drawers, plugging and unplugging connectors, pick‑and‑place, wiping tables, folding towels, and pouring medicine. (pandaily.com) The project README describes a “spatial semantic enhancement” that fuses textual instructions with 2D/3D spatial cues and full‑dynamics prediction to strengthen geometric understanding and manipulation generalization. (github.com) Model checkpoints and the accompanying G1 manipulation datasets were published on a Unitree Hugging Face collection and several per‑task datasets were updated on Jan 29 with viewers and downloads available. (huggingface.co) Unitree’s multi‑task model reportedly reaches near state‑of‑the‑art performance on the LIBERO simulation benchmark in their reporting, indicating competitive sim‑to‑real capability for multi‑task manipulation. (pandaily.com) The G1 platform Unitree targets for these experiments is marketed in base SKUs around the ~$16,000 range and ships in configurations with roughly 23–43 degrees of freedom and EDU variants that can include an NVIDIA Jetson Orin (100 TOPS) for onboard compute. (blog.robozaps.com)
