Pelican-VLA 0.5

Pelican-VLA 0.5 is the first vision-language-action model published by the Beijing Innovation Center of Humanoid Robotics — the same state-backed unit that builds the Tiangong humanoids and the Pelican-VL perception models. Where Pelican-VL describes a scene, this model outputs robot motion. The design point is a deliberate restriction. Between perception and action the network places 32 learnable bottleneck tokens, and the action pathway is barred by an attention mask from reading dense visual tokens directly. Everything the policy acts on has to pass through that fixed-capacity interface. The centre reports that this constraint alone produces attention focused on the object named in the instruction and on its contact region — without segmentation masks, object annotations or attention supervision, and before any task-specific tuning. The policy runs on a single Qwen3-VL 4B backbone (about 5 billion parameters in total), with a flow-matching action head producing chunks of 50 actions in ten denoising steps, and a frozen NVIDIA Cosmos tokenizer supplying visual history and future-frame prediction. Pre-training used roughly 2,400 hours of heterogeneous manipulation data across different robot bodies. On the RoboTwin 2.0 benchmark, after fine-tuning, the centre reports a 91.2% average success rate — the best figure in its table of open-source baselines, ahead of Hy-VLA (90.5) and JoyAI-RA (89.9) — with a gap of only 0.4 points between clean and randomised settings. The paper is unusually blunt about what is missing: the authors call the model an intermediate stage and name a remaining representation-to-action gap, meaning it begins to identify what to act upon before it reliably acts. The model card states plainly that this is a research checkpoint and not a production controller. Weights are Apache 2.0, but running them requires two further downloads under separate licences: the Qwen3-VL backbone and NVIDIA's Cosmos tokenizer.

▸Videos

No videos yet.

▸Reviews