Synced · WeChatJun 3, 2026, 15:30 UTC+8

RSS 2026: Ant Lingbo Proposes Autoregressive Causal World Model for Robot Manipulation with 50 Demos

RSS 2026｜蚂蚁灵波提出首个自回归因果世界模型，50条数据解锁通用机器人操控

Ant Lingbo and HKUST introduced LingBot-VA, an autoregressive video-action world model that unifies visual dynamics prediction and action inference, and the paper reports fine-tuning with 50 real-world demonstrations per task plus 92.0% and 91.1% success on RoboTwin 2.0 Easy and Hard settings.

Why it matters: HKR-H/K/R all pass: the hook is 50-demo robot control, with a concrete video-action world-model mechanism. Single-source coverage lacks code, benchmark detail, and deployment evidence, so it lands at 78.

Read the original ↗Export Markdown

Take

LingBot-VA’s strongest claim is not “first causal world model”; it is the ablation tying video prediction to control. The article gives a rare clean hook: removing video prediction drops success from 92.93% to 48.31%, and replacing causal modeling with bidirectional attention drops it to 81.46%. That makes the result harder to dismiss as VLA trajectory memorization.

The title still oversells “general manipulation.” LingBot-VA needs 50 real demonstrations per task for fine-tuning, while the 10-demo claim only says it beats baselines. Inference is 0.5 seconds per closed-loop step on one RTX 5880 Ada, so roughly 2Hz. That fits folding clothes or unpacking boxes. It does not fit fast, contact-heavy assembly. Beating π0.5 and Genie-Envisioner matters, but the article does not list per-task real-world success rates for the six physical tasks.

Source: Synced · WeChat · mp.weixin.qq.com
Published: Jun 3, 2026, 15:30 UTC+8 · Jun 3, 2026