Ant Lingbo embodied AI contest: 15 teams, one shared VLA model, 48-hour real-robot showdown
机器之心 · wechat · 2026-11-19
The first Ant Lingbo Embodied Foundation Model Challenge drew 3,000+ participants and 1,855 teams across 13 countries and 648 institutions. All teams trained the same open-source LingBot-VLA2.0 base model in RoboTwin2.0 simulation, then 15 finalists spent 48 hours deploying on real AgileX COBOT MAGIC dual-arm robots.
Team TARS (FuturLaboratory, NWPU, Pudu, Tianjin University) won first prize and 80,000 yuan. Notable technical routes: TARS used Qwen to detect distractors and re-collected data for a new dustpan-grasping grip; runner-up EVA built a DAgger data-flywheel, adding 700+ trajectories to lift real-robot success from 10% to 70%; SST used DeepSeek plus SAM3 with background dimming and reverse-video auto-labeling, adding only 13 samples; FDU-SONIC focused on data coverage.
A telling stat: 51.7% of entrants had VLA/multimodal backgrounds but only 20.9% had real-robot experience. Panelists flagged sim-to-real gaps (an oiled steel wheel dropping success from 95% to <10%), deployment cost as an evaluation criterion, and Ant Lingbo announced a LingBot developer ecosystem program targeting cross-embodiment adaptation and better benchmarks.
More from Embodied
- Antigravity Robotics plugin lets anyone build MuJoCo sim tasks with prompts — DynamicWebPaige · 2026-11-19
- META-Robot from HKUST completes 44/56 real-robot trials with self-repair — jiqizhixin · 2026-11-19
- Humanoids make sense for entertainment and high-risk human interaction, not factories, argues cixliv — cixliv · 2026-11-19
- Felis: MIT's Cat-Inspired Robot Uses an Actuated Spine With Just Six Motors — willknight · 2026-11-19
- Musk: AI and robotics are pulling SpaceX and Tesla together — rohanpaul_ai · 2026-11-19
- Musk: Neuralink's Blindsight implant aims to let completely blind people see — elonmusk · 2026-11-19