Agility Robotics Lines Up $465 Million Ahead of Listing, Scale AI and Hugging Face Open FineART, NVIDIA Releases Metatron

Nvdium Tech | October 30, 2026

Agility Robotics expects about $465 million as Nasdaq debut nears

Humanoid robot maker Agility Robotics and the SPAC Churchill Capital Corp XI said that, with the redemption deadline now passed, gross proceeds from their merger and related financings should come to about $465 million before transaction expenses:

• about $164 million from Churchill XI's trust account;
• about $201 million from a common-stock PIPE at $10 a share, led by Foxconn;
• $100 million in equity funding already received after the deal was announced.

Churchill XI shareholders vote on November 2. If approved, the deal is expected to close on November 4, and the combined company, still called Agility Robotics, should begin trading on Nasdaq under the ticker AGLT on November 5. On the commercial side, Agility has expanded its work with GXO to two more facilities, signed AGCO for a manufacturing pilot due to start in 2027, and has three more manufacturers in paid programs. Digit 5, unveiled in September, is set to reach early-access customers in the first half of 2027, with general availability targeted for the end of 2027. Agility has previously disclosed more than $300 million in committed multi-year orders for Digit 5.

Scale AI and Hugging Face open-source the FineART dataset and FineART-VLA

Scale AI and Hugging Face's LeRobot team released FineART, the largest open robot dataset with subtask labels. It holds 40,543 complete task episodes, or 1,718 hours of footage across 151 tasks, with 533,913 labeled subtask steps, about 13 per episode. Pick-and-place and sorting make up 77% of the data; the other 23% covers harder work such as textiles, cables, tool use and precision insertion. Separate teams handled collection, subtask annotation and quality checks. Alongside it came FineART-VLA, the first open System 1–System 2 VLA in LeRobot: it first predicts the next subtask in language, then turns it into bimanual actions. Across 3,400 real-robot evaluations, the model improved at spatial instructions such as left, right and middle, at step-by-step long-horizon tasks, and at transfer to new hardware. It is Scale AI's first open-source robotics dataset and model.

NVIDIA open-sources Metatron v1, one policy for whole-body control of Unitree G1

NVIDIA released Metatron v1 on Hugging Face, a single frozen policy that controls the 29-joint Unitree G1 humanoid for walking, scene interaction and getting up after a fall. Instead of a full pose at every instant, Metatron takes rough goals for the hands, feet and torso, such as position, orientation, velocity and arrival time, and fills in the rest of the body's motion. New skills come from code that sets those goals, while the policy weights stay fixed: a coding agent writes the goal code, tests and refines it in simulation, and a person decides whether to deploy it on the robot. Training happens in two stages, first a reinforcement-learning tracker for dense whole-body motion, then masked on-policy distillation into Metatron. In simulation, the same policy clears four hurdles in one run, falls into a bar and gets back up to keep running, bowls, hits a tennis forehand and kicks a ball. On hardware, it shows gamepad-steered walking, teleoperation, and walking while being guided by hand. The model is integrated into the ProtoMotions framework and ships ONNX versions that run in a web browser.