독서리서치레이더투자 프레임워크
로그인 / 회원가입
← 읽기로 돌아가기
뉴스주제 · 蚂蚁灵波VLA模型중대3 분 소요DisconfirmAI 편집부

First Ant Lingbo Embodied AI Challenge Concludes: Developers Showcase New Synergy Between General LLMs and Embodied Foundation Models

The inaugural Ant Lingbo Embodied AI Challenge concluded in Shanghai, with winning team TARS scoring 82 points. The competition demonstrated that leveraging data engineering and integrating general LLMs like DeepSeek for task understanding significantly enhances LingBot-VLA 2.0's performance in real-world scenarios, though the gap between simulation and reality remains a challenge.

2026-11-17 19:18 GMT+8

On November 16, the first Embodied AI Challenge hosted by Ant Lingbo concluded in Zhangjiang, Shanghai. After online preliminaries, 15 teams advanced to a 48-hour offline final. All teams used the same open-source Vision-Language-Action (VLA) model, LingBot-VLA 2.0, along with the Songling COBOT MAGIC robot and unified on-site computing power, to complete three real-world physical tasks: vegetable and fruit placement, cosmetics organization, and dual-arm cleaning.

The competition results revealed a key trend in current embodied AI development: success no longer depends solely on who collects more data, but on how developers cleverly utilize the capabilities of existing foundation models. The winning team, TARS, achieved an 82-point score by prioritizing task difficulty and introducing large models to identify target objects and distractors. By occluding irrelevant areas, they allowed the robot to focus on the current task, achieving a 100% success rate in the vegetable and fruit placement task. Additionally, TARS compressed single-step training time from approximately 15 seconds to 3.4 seconds, gaining valuable time for rapid on-site iteration.

The third-place team, SST, demonstrated a "less but precise" strategy. They collected only 13 additional data points but integrated DeepSeek during inference to understand task instructions and used visual segmentation models to locate target objects before handing off actions to LingBot-VLA 2.0, ultimately scoring 65.87 points. This approach concretely illustrates the synergy between general-purpose large models and embodied foundation models: DeepSeek solves the understanding problem of "what to do," while LingBot-VLA 2.0 handles the execution problem of "how to do it." SST also froze some pre-trained parameters to avoid disrupting the model's original general capabilities with small amounts of supplementary data.

로그인 / 회원가입
독서리서치레이더투자 프레임워크

Other teams adopted different technical routes. Second-place EVA chose continuous data augmentation, converting failure cases into training data and improving their self-tested success rate from about 10% to 73.83% after seven iterations. The resource-limited FDU_SONIC team deeply analyzed the 647 trajectories provided by the organizers, adjusting data based on real-world test failures, and secured fifth place. These differences indicate that as foundation models become stronger, developers need less redundant construction of underlying capabilities and can focus on solving specific real-world problems.

However, the competition also exposed realistic challenges facing embodied AI. The team "Today Also Check Things Well," coming from the field of large model evaluation, found that although they validated their solutions in simulation environments, a discrepancy of about 5 centimeters between simulated and real tabletops caused their designed actions to fail on-site. This directly demonstrates that precise reproduction in digital environments cannot automatically bridge the distance between simulation and reality; robots still face uncertainties such as lighting, friction, and mechanical errors in the physical world.

At the concurrent Ant Lingbo First Developer Conference, CEO Zhu Xing stated that embodied intelligence is still in its early stages. He noted that progress in general large models accelerates industry development, but robots must ultimately prove their adaptability, speed, and cost feasibility in real-world scenarios. He pointed out that the industry does not yet know which type of data scaling will lead to a capability leap, meaning embodied AI has not yet found its own "Scaling Law." The significance of this competition lies in enabling developers to experiment around the same foundation model, bringing back successful methods, lessons from failures, and real-world problems, which is valuable for driving the convergence of technical paradigms.

출처

  • geekpark.netGeekPark↗
주제

蚂蚁灵波VLA模型

蚂蚁集团旗下灵波科技(Robbyant)推出的具身智能基座模型 LingBot-VLA 系列,正通过“一脑多机”策略推动机器人规模化落地。2026年7月,该公司全面开源升级后的 LingBot-VLA 2.0,该模型基于6万小时真实物理数据…

3 단계1 건업데이트됨 2026-11-20 07:04

현재까지의 과정

  1. 2026-01-28蚂蚁灵波宣布全面开源首款具身大模型 LingBot-VLA 1.0,提供模型权重及全套代码库,旨在降低商业化落地门槛。finance.sina.com.cn ↗
  2. 2026-07-08蚂蚁灵波正式开源 LingBot-VLA 2.0。该版本预训练数据量增至6万小时,覆盖17家厂商的20种机器人构型,支持头部、腰部及移动底盘等自由度,推理延迟控制在130毫秒以内。21jingji.com ↗
  3. 2026-11-17首届蚂蚁灵波具身大模型挑战赛在上海收官。冠军团队 TARS 得分82分,比赛显示接入 DeepSeek 等通用大模型辅助理解任务能显著提升 LingBot-VLA 2.0 在真实场景中的表现。geekpark.net ↗

아직 해결되지 않은 문제

  • 通用大模型与具身基础模型的协同是否已形成可复制的标准范式,还是仍依赖特定团队的工程技巧?
  • LingBot-VLA 2.0 在商业试点(如零售分拣、物流)中的实际部署成本和效率提升幅度如何?
  • 行业尚未收敛的技术路线中,哪类数据扩大规模后能带来真正的能力跃迁?
전체 주제 읽기 →
리서치 계속하기 →