China Hot Tech @ChinaHotTech
← All posts
Embodied · Open Source · Robotics

Ant LingBot / Robbyant: Open Robot “General Brain” Hits Wuzhen 2026

蚂蚁灵波 (brand Robbyant), Ant Group’s embodied-AI company, brought a full-stack open LingBot “general brain” to the 2026 World Internet Conference in Wuzhen (Nov 2–5, theme 开源开放) — spatial perception, dexterous manip, and world/action models meant to reuse one brain across many robot bodies*, plus a live 国大药房 pharmacy aisle demo*.

· ~8 min read · China Hot Tech

Official Robbyant product still — white/black humanoid robot with Robbyant chest panel in a modern lobby

Real official product still from technology.robbyant.com (Robbyant / 蚂蚁灵波) — humanoid with “Robbyant” chest panel; not AI art.

Most robot teams still train a new policy every time the arm, gripper, or aisle layout changes. Robbyant’s Wuzhen pitch is the opposite: ship an open embodied stack that covers seeing space, predicting how the world moves, and emitting actions — then let labs and shops post-train the same “brain” onto different bodies*. LeiPhone and 封面新闻 both frame the booth around that “one brain, many robots” line, and the company brand on the hardware chest plate is simply Robbyant.

What showed up in Wuzhen

Press coverage of the 互联网之光 booth walks through a LingBot 2.0 matrix*: LingBot-Vision / LingBot-Depth 2.0 for spatial perception; LingBot-Video / LingBot-World 2.0 for how scenes evolve; LingBot-VLA 2.0 / LingBot-VA 2.0 for instruction→action and interleaved world-action control*. The copy is bilingual and marketing-heavy — treat the slogans as vendor framing — but the underlying repos are real Apache-licensed codebases under the Robbyant GitHub org, with weights on Hugging Face / ModelScope.

Vendor stats for the series after roughly nine months open*: ~2.2 M downloads*, >33k GitHub Stars*, and citations by Physical Intelligence and Generalist AI*. The same brief says Stanford OpenWAM reused LingBot-VA*, Korea FastOPD used LingBot-VLA as a teacher*, and EPFL used LingBot-Map for coral-reef 3D*. Asterisk every reuse claim until you read the papers yourself.

Can you use it?

Yes — real downloadable weights and code, not a booth-only teaser. Start at github.com/Robbyant/lingbot-vla (Apache-2.0) and huggingface.co/robbyant/lingbot-vla-4b; VLA 2.0 / VA / World live in sibling repos (lingbot-vla-v2, lingbot-va, lingbot-world / lingbot-world-v2). LeRobot users: Hugging Face also hosts a convert path under lerobot/lingbot_vla_v2_base*. Expect GPU-sized checkpoints, post-train glue, and robot-specific calibration — this is not a one-click phone app, and the Wuzhen pharmacy aisle is a deployed demo path*, not something you recreate from weights alone tonight.

Official LingBot-VLA Hugging Face Teaser — architecture diagram plus six real dual-arm robot demo tasks

Real media from Hugging Face robbyant/lingbot-vla-4b assets/Teaser.png: LingBot-VLA pipeline + six real dual-arm demos (Luban lock, cubes, lemon peel, scarf, sieve, towel fold) — not AI art.

Toolchain hooks that matter to builders

Beyond the model cards, the Wuzhen brief leans on ecosystem glue*: LingBot-VLA / VA adapted to LeRobot*; LingBot-Vision merged into timm mainline*; LingBot-World claimed as the first world model with both vLLM and SGLang* support; NVIDIA’s FlashDreams integrating LingBot-World*. Those are the kinds of hooks that decide whether a China open release shows up in your training loop or dies as a PDF. Verify each adapter in the current repo docs — release notes move faster than conference banners.

Community scale claim from the same brief: first Ant LingBot embodied-model challenge drew 13 countries, 648 orgs, 1855 teams, 3252 contestants*. Treat it as a press number for “people actually forked this,” not as a leaderboard you can reproduce.

Pharmacy aisle: the physical demo

At the Wuzhen tech hall, Robbyant partnered with 国大药房 on a smart-pharmacy scene*: dense shelves, an ~80 cm aisle, robot running order→pick→deliver without remodeling the store layout*. Press says the same path already runs in multiple Shanghai Guoda stores*. Industrial partners named for pick/place and logistics include 隆盛唯睿 / 德马科技 / 德迩象网*; body vendors 乐聚 and 钛虎 for multi-morphology checks*; plus an AFDE MoU aimed at Middle East retail/logistics/hospitality*. Awards on the booth: 互联网之光 新光 product (full-stack brain) and 新耀 scene (smart pharmacy)*.

What changes (and what does not)

Bottom line: if you have been waiting for a China embodied release that lands in the same folder as LeRobot and HF instead of only on a conference LED wall, Robbyant’s LingBot stack is the concrete watch item from Wuzhen 2026 — open VLA/World code*, a pharmacy aisle that already has store footprints*, and a long list of vendor asterisks worth reading before you budget a post-train run.

Sources