Unitree open-sources UnifoLM-WLA-1.0 humanoid foundation model (Apache-2.0)

Description by Gemini, which watched the video:

Here is the catalog entry for the video:

Summary
This official announcement video from Unitree Robotics showcases the major open-source release of UnifoLM-WLA-1.0, a general-purpose foundation model for humanoid robots. The video presents benchmark evaluation results comparing UnifoLM against leading vision-language and embodied AI models, followed by extensive demonstrations of autonomous whole-body manipulation and household chores running on a Unitree humanoid robot.

What is shown

• [00:00 - 00:01]: Title title card: "Fully Open Source UnifoLM-WLA-1.0: Unitree General-Purpose Humanoid Foundation Model Fully Upgrade Major Open Source".
• [00:02 - 00:04]: Benchmark comparison tables showing "Embodied Reasoning Benchmark Results" (evaluating RoboBrain, Helix, Qwen2-VL, Gemini, GPT-4o, etc., on benchmarks like RoboVQA, Ref, Where2Place, Pix2Point, Spatial Understanding, BLINK, VSR, and Multimodal Understanding).
• [00:05 - 00:27]: A 2× speed multi-panel montage showing "Autonomous Execution" of dozens of dexterous tabletop tasks: inserting screwdrivers into toolboxes, flipping books, placing plates into dish racks, packing boxes with tape, sorting parts into bins, wiping surfaces with cloths, pouring, peg-in-hole manipulation, handling flexible fabrics, and arranging flowers.
• [00:28 - 01:13]: Real-time footage (captioned "Real Footage Throughout No Speed-Up Autonomous Execution") displaying real-time head/wrist camera feeds and terminal telemetry (execution step action arrays, policy latency ~100–108 ms). The humanoid squats, picks up a laundry basket from a table, walks across the room, sets it on a chair, opens a front-loading washing machine door, and loads laundry into the drum.
• [01:14 - 01:35]: The humanoid robot picks up a plastic bottle from a low side table, lifts a tied plastic garbage bag out of a small wastebasket, walks over to a tall yellow wheelie bin, opens the hinged lid with one hand, drops the bag inside, and lets the lid close.
• [01:36 - 01:58]: Kitchen manipulation: the robot carries a mug across a kitchen, pulls open a lower dishwasher drawer, picks up a pink dish from the counter, places it into the rack, and slides the drawer closed.
• [01:59 - 02:16]: Shoe rack organization: the robot approaches a shelf, bends down, picks up a slipper, places it neatly onto a shoe shelf, and aligns it.
• [02:17 - 02:43]: Bathroom cleaning and grooming: the robot straightens a hanging pink hand towel on a towel bar, taps a wall-mounted mirror control panel, picks up a tube of toothpaste from the sink counter, and places it neatly inside a cup.
• [02:47 - 02:49]: Unitree disclaimer card advising customer safety distances (at least 2–3 meters) and noting ongoing research exploration in humanoid robotics.

Claims & numbers

• Open Source: The title and opening slide declare UnifoLM-WLA-1.0 to be a "Fully Open Source" general-purpose humanoid foundation model.
• Benchmark Performance: The benchmark table shows UnifoLM-ER-1.4B achieving scores of 62.7 on RoboVQA, 54.2 on Spatial VSR, 58.1 on Real World, 2320.8 on MMMU Val, and top scores across several BLINK/Pix2Point spatial understanding categories compared to models like RoboBrain2.0-7B, Helix-7B, and Qwen2-VL-7B.
• Execution Speed: The tabletop tasks are explicitly marked as "2×Speed Autonomous Execution", while the continuous whole-body household tasks are labeled "Real Footage Throughout No Speed-Up Autonomous Execution".
• Inference Latency: The terminal HUD indicates real-time policy inference running at ~100–108 ms latency per cycle.

Notable quotes

• [00:00]: "Fully Open Source UnifoLM-WLA-1.0 Unitree General-Purpose Humanoid Foundation Model"
• [00:28]: "Real Footage Throughout No Speed-Up Autonomous Execution"
• [02:47]: "Currently, the humanoid robot field is in the early stages of exploration worldwide."

Assessment
This is an official demonstration video from Unitree Robotics validating their open-source UnifoLM-WLA-1.0 model on physical hardware. The demonstrations showcase genuine autonomous execution with real-time multi-camera telemetry and policy outputs shown on-screen, though the initial multi-task montage is presented at 2× playback speed as disclosed.

Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.