---
id: "lib-2026-11-08-after-gpt-6-embodied-foundation-models-box"
type: "library"
title: "After GPT-6, embodied foundation models box themselves in"
title_original: "GPT-6之后，具身大模型画地为牢"
lang: "ZH"
org: "量子位 QbitAI (on 36Kr)"
source_type: "Practitioner post"
date: "2026-11-08"
week: "2026-W45"
url: "https://nosov.ai/library/lib-2026-11-08-after-gpt-6-embodied-foundation-models-box/"
original_url: "https://www.36kr.com/p/4060707751826311"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "QbitAI trade feature built on interviews with the first author of the RoboDojo benchmark and several robotics founders; useful industry context, but the score gains are simulation-only and the job-cut reports are unattributed."
summary_basis: "full-text"
original_minutes: "9"
tags: ["ai technology", "embodied ai", "independent evaluation", "labour"]
frames: ["verification-capacity"]
---

# After GPT-6, embodied foundation models box themselves in

Original title: GPT-6之后，具身大模型画地为牢

Tier B (Useful, limited): QbitAI trade feature built on interviews with the first author of the RoboDojo benchmark and several robotics founders; useful industry context, but the score gains are simulation-only and the job-cut reports are unattributed.

Summary based on the full text.

## In short

A 8 November 2026 feature by QbitAI (量子位), carried by 36Kr, on how OpenAI's GPT-6 Astra unsettled Chinese embodied-AI companies that train their own robot foundation models. In the RoboDojo evaluation, a general model without task-specific post-training or demonstrations handled in-context learning, out-of-distribution correction and open-vocabulary manipulation, three problems embodied start-ups had built pitches around. Real-robot tests were halted for safety and API latency is high, so the article ends on economics: use rules or small models where tasks are stable and general models where tasks keep changing.

## Key numbers

- 20-30 to 50-60: RoboDojo scores of teams' models after combining them with GPT-6 Astra-based approaches
- 80 to 90: interviewee's analogy: in-house embodied models add 0.1-0.2 points at 80, Astra jumped to 90 at once
- about 2 months: time since GPT-6 Astra's release, with GPT-6.2 Astra expected soon

## What it reports

- Industry reports say some embodied-AI companies have started cutting algorithm jobs; no companies are named.
- Per Chen Tianxing (HKU MMLab, first author of RoboTwin/RoboDojo), teams that had trained native embodied models now submit solutions using general models via Robot-Use-Agent; scores rose from 20-30 to 50-60.
- Astra results: image or text demonstrations raised success over none; it adapted after deliberate camera and coordinate changes; it re-planned a grasp when an object moved; it followed varied phrasing without task post-training.
- The RoboDojo report states Astra's real-robot tests were stopped over safety, and simulation scores do not equal real-world productivity.
- Weaknesses named: continuous contact, multi-step physical manipulation, force and deformation (per founders Xu Huazhe and Wang Qian), and high cloud API latency for closed-loop control. An engineer notes slow inference can still allow fast low-level control for plannable tasks such as parcel sorting.
- Chen's framing: open tasks lean on general models, stable tasks on rules or light policies, the middle on scene-optimised models. OpenAI has been hiring heavily for robotics roles this year.
- Mood: "investors have stopped investing in embodied AI", which the article itself calls possibly exaggerated.

## How they got it

Journalistic piece drawing on the RoboDojo team's early evaluation report and interviews with its first author, two founders and anonymous engineers.

## Limits

- Benchmark results are zero-shot simulation (RoboDojo-Sim) only; no real-robot numbers.
- Job cuts and investor sentiment are hearsay with no names or figures.
- Trade-media tone; score gains are quoted without the full table.

## Why it matters for companies

For buyers of robotics the useful point is the cost test: on a fixed production line, success rate, takt, stability and price per call matter more than general intelligence, so a cheap specialised model may beat an expensive general one. For anyone funding in-house foundation models the story is a warning that a general model can absorb years of specialised work in one release; investment should go where general models remain weak, here physical contact, hardware control and deployment.

## Key passages in translation

> 具身自研模型还在80分上不断加0.1、加0.2，GPT-6 Astra已经一次性从80分跳到了90分。

In-house embodied models keep adding 0.1 or 0.2 at 80 points; GPT-6 Astra jumped from 80 to 90 in one go.

> Astra的真机测试曾因安全问题中止。

Astra's real-robot tests were once halted over safety problems.

> 如果花几毛钱调用一个模型就能稳定完成任务，为什么非要花几块钱调用GPT-6 Astra？

If a model costing a few cents per call can do the task reliably, why pay a few yuan to call GPT-6 Astra?

Original: https://www.36kr.com/p/4060707751826311