Unitree releases its humanoid base model and fine-tuning code

Unitree’s Chinese-language project README lists the September 28 release of UnifoLM-WLA-1.0-Base and fine-tuning code. It follows September releases of perception-model weights and code for training action experts.

Unitree describes UnifoLM-WLA-1.0 as a six-billion-parameter humanoid foundation model trained using approximately 2,500 hours of real-robot data, with a single model covering 64 tabletop and whole-body manipulation tasks. These task and training figures are the company’s reported results, rather than independent benchmarks reproduced by SuperRobots.