---
id: "lib-2026-10-09-tsinghua-linked-embodied-model-tops-global"
type: "library"
title: "Tsinghua-linked embodied model tops global leaderboard (VPP2)"
title_original: "清华具身模型登顶全球第一！突围GPT-6、英伟达，不靠外挂和额外数据"
lang: "ZH"
org: "量子位 QbitAI"
source_type: "Release"
date: "2026-10-09"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-09-tsinghua-linked-embodied-model-tops-global/"
original_url: "https://www.qbitai.com/2026/10/502125.html"
tier: "C"
tier_name: "Weak signal"
tier_reason: "Promotional trade-press write-up of a company's own paper and leaderboard result; the numbers are specific but the article adds no independent assessment."
summary_basis: "full-text"
original_minutes: "10"
tags: ["ai technology", "embodied ai", "world action models", "benchmarks"]
frames: ["verification-capacity"]
---

# Tsinghua-linked embodied model tops global leaderboard (VPP2)

Original title: 清华具身模型登顶全球第一！突围GPT-6、英伟达，不靠外挂和额外数据

Tier C (Weak signal): Promotional trade-press write-up of a company's own paper and leaderboard result; the numbers are specific but the article adds no independent assessment.

Summary based on the full text.

## In short

A QbitAI (量子位) article by Tian Yanlin, dated 9 October 2026, about VPP2, a "world action model" from Robotera (星动纪元), a robotics company in which Tsinghua University holds a stake. VPP2 ranks first on RoboDojo, a simulation benchmark for robot manipulation led by HKU MMLab with nearly 20 institutions. The method trains video prediction first and action generation separately, so the model's ability to predict physical change on new tasks is not lost when it learns to act. The article states the result used only the standard dataset, without extra data or enhancement tricks.

## Key numbers

- 32.26%: VPP2 average success rate on the RoboDojo simulation leaderboard, vs 22.48% for GPT-6-Astra
- 58.5% vs 40%: zero-shot success on 10 real-robot ALOHA task types, VPP2 vs π0.5
- 27.6% to 57.6%: long-horizon success rate after adding a VLM sub-task planner
- 0.22 s: latency to generate an action chunk (0.12 s video prediction plus 0.1 s action expert)

## What it reports

- RoboDojo: 42 dual-arm tasks across five dimensions (generalisation, precise manipulation, long-horizon, memory, open-vocabulary instructions).
- VPP2 average success rate 32.26% and average score 39.26, ahead of GPT-6-Astra (22.48%, 28.97) by 9.78 points and 10.29 points; first in generalisation, precision and memory.
- Real ALOHA robot, 10 zero-shot task types: 58.5% average success vs 40% for π0.5, best on 9 of 10.
- Built on Alibaba's open Wan2.1-I2V-14B; the 14B model reached 90% instruction-following success in robot video prediction vs 78% for the 64B Cosmos3.
- Three-stage training: event-level video pre-training, fixed 8-second clips with consistency distillation, then a 0.9B action expert trained with the video model frozen and adapted by LoRA.
- LIBERO-Pro overall success 45.0% vs at most 11.0% for baselines; LIBERO-OOD 63.9%.
- Adding a VLM planner raised long-horizon success from 27.6% to 57.6%.
- The company works with China Post and SF Express in more than 10 logistics centres across 5 provinces and cities; the article notes this is not yet scaled deployment of VPP2. Code is open source.

## How they got it

The article summarises the company's paper and project page; baseline figures are taken from the official leaderboard. No independent replication is reported.

## Limits

- Even the leading score means about two in three simulated tasks fail.
- Simulation results and a 10-task real-robot test are small evidence for warehouse use; the logistics partnerships are not linked to VPP2 performance data.
- The text reads as promotion and gives no criticism of the method.

## Why it matters for companies

For operations and logistics managers, the numbers set expectations: the best general robot manipulation model succeeds on about a third of hard simulated tasks and just over half of simple real ones. The pairing of a language-model planner with a physical executor doubled long-task success, which points to where gains are coming from. Pilots in fixed, specialised settings remain the realistic path for now.

## Key passages in translation

> 星动纪元选择将视频预测与动作学习分阶段训练，重点不是「视频、动作一锅炖」，而是把两者「解耦」，重新「排序」。

Robotera chose to train video prediction and action learning in separate stages; the point is to "decouple" the two and "re-order" them instead of "stewing video and action in one pot".

> VPP2提出了一个很直接的判断：视频预测的质量，决定了动作的上限。

VPP2 puts forward a very direct judgement: the quality of video prediction sets the upper limit of action.

> 在机器人操作视频的指令遵循测试中，14B参数的VPP2达到90%的成功率，而64B参数的Cosmos3模型为78%。

In an instruction-following test on robot manipulation videos, the 14B-parameter VPP2 reached a 90% success rate, against 78% for the 64B-parameter Cosmos3 model.

> 在RoboDojo选定的长程任务测试中，引入VLM子任务规划后，平均成功率从27.6%提升至57.6%。

In selected RoboDojo long-horizon tasks, introducing VLM sub-task planning raised the average success rate from 27.6% to 57.6%.

> 当然，已有物流业务的商业化进展，并不意味着VPP2已经实现规模化部署。

Of course, commercial progress in existing logistics business does not mean VPP2 has achieved scaled deployment.

Original: https://www.qbitai.com/2026/10/502125.html