One bowl of stacked dishes costs 19 USD: SenseTime open-sourced an embodied framework riding the GPT-6 Astra hype. How did it get 450 stars

The SenseNova team wrote an evaluation report asking "does GPT-6 Astra have embodied intelligence?" and open-sourced a framework called SenseNova-RoboRSI, packaging a dozen-plus robot simulation leaderboards into Docker, evaluable with one command. It got 450 stars in 28 days, with the peak falling on the day after the paper hit #1 on Hugging Face and Chinese finance media ran coverage together. The stars came mainly from the report's topic and big-company PR, which followers can't borrow; the transferable part is the playbook: "run a question-titled evaluation on the hottest model, then open-source a framework to catch the traffic."

One bowl of stacked dishes, 131 minutes

The most memorable numbers in the report are clearest in the long piece on 36氪 11-06: having GPT-6 Astra directly control a robot to stack one bowl takes 131 minutes on average, with inference costs around 19 USD; putting a pen into a pen holder averages 263 minutes and about 55 USD.

The report's conclusion has two halves: "Embodied capability is largely there; practical embodied intelligence is not yet." With steps relaxed to 10× and the model allowed to look at the training set first, Astra nears perfect scores on simulation leaderboards like LIBERO and SIMPLE, and RoboDojo success reaches 81.9% (only 27.6% at standard step counts). The cost is slowness and expense: one LIBERO-Long++ episode takes roughly 1M tokens and nearly an hour.

SenseTime's answer: on the same task, using its open-source framework with its own 27B model cuts the cost to 1/50 of what calling Astra directly costs. That "1/50" line was then quoted nearly word for word by almost every outlet. 3636Kr also pointed out that SenseTime ran the test itself, only on specific tasks like stacking bowls and putting away pens, and no third party has re-run it yet.

What it is

A runtime for using LLMs as robot brains, which its author calls Harness. Models, prompts, skills, and tools (including MCP) plug in, and it runs and scores on different simulation leaderboards under each board's own rules. Each leaderboard and its dependencies are packed into a Docker image behind one unified HTTP interface. Write one config file and you can run an end-to-end evaluation, without installing each simulator by hand. Anyone who has done embodied research knows environment setup is the biggest time sink.

On top of that sits RSI (recursive self-improvement): it uses execution traces and failure feedback to modify the Harness itself, such as how planning is organized, how feedback is handled, and how memory is stored, without touching model weights. Each change is re-tested under the same conditions, and only the ones that work move to the next version. The validated experience is then used to train a 27B model (based on Qwen3.8-27B, with 10-22 releasing the weights, downloadable from Hugging Face as 317,11-06).

Python, Apache-2.0. The README itself labels it Pre-Alpha. The repo description still says "Code release coming soon," even though the code for each leaderboard has already been released in five batches.

When the stars came

Data: GitHub daily stars / forks (from 10-17), and the update log in the README.

You can see two phases. The first is a spike: #1 on the paper page plus press releases pushed it up about 120 in three days. The second is slow growth: code gets released batch by batch, and every three or four days the README gains a new "leaderboard code" or "#1 on a leaderboard" entry, with stars creeping up behind it. Forks went from 10-27 to 10 to 11-06 to 33, mostly after code releases, which suggests people are starting to actually run it.

Searching the name on X turns up almost no discussion in the last 30 days, and I found no dedicated Hacker News or Reddit threads. It spread through small embodied-AI research circles, and big-company PR pushed it into Chinese media.

Why this one

It answered a question the industry is already arguing about. OpenAI released GPT-6 Astra in September, and the next day someone tested it controlling a real robot arm. "Can an LLM act directly as a robot policy?" has been debated for a month. At the same time, Awesome-Astra-Embodied-AI, a plain resource list, got 1,234 stars, far more than this framework. The topic itself draws more traffic than any single framework.

The report got seen before the repo did. The report title is a question, and its conclusion is "half yes, half no," which makes it easy to repost. The Hugging Face paper page links straight to GitHub, and the day it hit #1 on papers was also the day it got the most stars.

A listed company's PR machine. On 10-20, a batch of finance outlets published the same story at once, with the same framing: 12 simulated charts, 8 real devices, cost 1/50. 36 Pinwheel's version pushed the angle that SenseTime had just delivered its first profit report since going public, and it needed a new story. Indie devs don't have that kind of resource.

One release split into eleven news items. From 10-09 to 11-03 the README update log has eleven entries: five code batches, an image batch, and an experience pack, interspersed with "#1 on a leaderboard" and "weights released." Each one gives researchers a reason to click back in.

There's a discount to apply too. allagent.wiki's 10-21 breakdown points out that the RoboDojo 60.08 points belong to the combination of "SenseTime's Harness + GPT-6 Astra," not to SenseTime's own model. At the time RoboDojo's integration code had not been released, so others couldn't reproduce it, and the 27B model on its own scored only 20.83%. Comparing Harness alone, on LIBERO-PRO it leads a similar approach, RPent, by only 1.87 points (94.50% vs 92.63%).

Who else is doing the same thing

I haven't seen a repo built specifically around this yet. The closest concept, ros-claw/PhysicalEvoBench, has 0 stars.

Can you follow it?

First, strip out what can't be borrowed: SenseTime's PR, the team and professor co-authors behind the Hugging Face #1 paper, and the exposure from the org's two repos with over 6,000 stars. Those together account for most of the 450 stars. Even so, 450 isn't a lot, which suggests the embodied-framework circle is small and the people starring it are mainly researchers.

Three things are worth learning, and they apply well beyond embodied AI:

1. Run a serious evaluation of the hottest closed-source model, title it as a question, and open-source the eval code. The report gets shared, and the repo catches the traffic. In the weeks after a new model launches, everyone wants to know "can it actually do X?"
2. Package hard-to-install benchmarks into Docker with a unified interface. This is the most practical value in RoboRSI, and one person can do it. Other fields have the same pain point: a pile of benchmarks, each with its own setup.
3. Release code in batches, with one update post per batch. A single big release gets one day of attention. Splitting it into eleven updates gives eleven reasons to pay attention.

The places it falls short are also opportunities: its headline results can't be reproduced, there's no independent re-testing, it's mostly simulation, and the cost numbers only cover a few tasks. An independent, reproducible head-to-head evaluation of "who can actually control robots" doesn't exist yet. But that takes compute and robots, so the entry bar is not low (my inference).

If it were you

C: worth reading to understand its playbook, but building an embodied framework to chase stars isn't worth it. Its stars came mainly from the report's topic and SenseTime's PR, not from the framework itself. The embodied circle is small, and big-company backing plus press releases only got it 450 stars. Under the same topic, a resource list got nearly three times as many stars, which shows the traffic lives in the topic, not the framework.

• Who it fits: grad students or engineers who work in a specific research area, have compute and evaluation experience, and can produce a solid evaluation within two to three weeks of a new model launch. Or people who understand the "benchmarks are hard to install" pain point in some field.
• Who it doesn't fit: anyone who wants to build yet another embodied Harness head-on (several already exist in the same wave, including RPent and Embodied-Harness, all backed by labs), or anyone without robotics or simulation experience who just wants to ride the hype.
• How to do it: watch the next major model release, pick a capability people are arguing about that no one has systematically tested, publish the evaluation first, then open-source the code. Package the eval code in Docker with a unified interface, and put "run everything with one command" on the first screen of the README. Publish a report with a question as the title at the same time. For less effort, start with a topic list (an Awesome-style repo), which got the most stars in this wave.
• How to get users: submit the report to Hugging Face Daily Papers and arXiv, and link the paper page to the repo. Find an influential professor or researcher in the field to co-author and help repost. Then release code in batches, posting one update each on the README and X.
• How long the window lasts: the Astra topic has already been hot for two months, so the window in embodied AI is probably only a few weeks. This playbook reopens a window with every major model release, though.
• Signs I'm wrong: (1) RoboRSI gained 3,000 stars in a month, which means the GitHub audience for embodied frameworks is bigger than I assumed, and a standalone reproduction-evaluation or multi-benchmark packaging tool would be worth a step up. (2) Someone without big-company backing gets over a thousand stars on this topic with just one evaluation and open code, which means the playbook matters more than resources. (3) RoboDojo's official leaderboard lists its scores and third-party re-tests match, so it could become the standard tool in this area. Then the better move is to write plugins for it and integrate new benchmarks, rather than starting a competing project.