Staff cuts are reaching algorithm teams at Chinese embodied AI companies, and the trigger is not a lack of capital but a general-purpose model that performs robot tasks without being trained for them.
What changed in two months
GPT-6 Astra shipped two months ago and a further version is expected shortly. The embodied AI industry has not recovered from the first release.
Engineers working on pretraining described the effect as immediate. Capabilities that previously required large volumes of data, compute and specialist engineers to grind out one task at a time were appearing in some tests without task-specific post-training at all.
Several embodied AI companies have begun cutting algorithm positions in response.
A model developer at a robot company told QuantumBit that the real shock is not the absence of a paradigm breakthrough inside embodied AI. It is that a general model is attacking the sector vertically, through coding, and doing it faster than the robotics industry can respond.
The pressure is not only from California. Chinese embodied companies are watching domestic model labs more closely than they watch OpenAI, and the names that now matter to them are Kimi, DeepSeek and Zhipu.
The benchmark that forced the question
In the RoboDojo benchmark, a growing number of teams have begun routing general models through their robot systems, with strong generalisation and unexpectedly high scores.
Chen Tianxing, a PhD researcher in the embodied AI group at HKU MMLab and first author of the RoboDojo benchmark, said teams that had insisted on training native embodied foundation models are now submitting approaches that combine a general model with a robot-use agent.
The effect is direct. Models that scored 20 to 30 points moved to 50 to 60 when paired with Astra-based methods.
The score gains raise a commercial question for every company in the sector, which is whether an in-house model is still worth the spend.
Three capabilities that were supposed to be hard
Chen described the jump with a comparison: native embodied models were adding 0.1 and 0.2 points on top of 80, while Astra moved from 80 to 90 in a single step.
In September, when his colleagues began bringing him test results, he said the impact on 3D reconstruction and mathematics was unlike anything he had seen.
The first surprise is in-context learning. A robot can learn a new task from a short amount of context, for example a video of a person performing an operation plus a text description, and then attempt to reproduce it with an arm. In July this was an active research direction in embodied AI, usually framed as one-shot imitation. Researchers found that supplying image or text demonstrations to Astra raised success rates above the no-demonstration baseline.
The online adaptation went further. When researchers deliberately changed the camera view and the action coordinate frame, Astra adjusted its strategy over a few attempts and completed the task. It does not only read context, it corrects while acting.
Chen’s explanation is that a general model already handles context across a conversation, and the same mechanism transfers when the context is a video and a text instruction.
The second is out-of-distribution recovery. In real environments a robot meets situations outside its training data. If an arm is reaching for a cup and the cup is knocked to a new position, a model reliant on the training distribution may continue on the original trajectory and grasp nothing. Astra re-plans from the changed visual input and retries the grasp.
The third is open-vocabulary manipulation, which concerns aligning language with action across objects that were never in a fixed label set.
What is left to defend
If a general model can drive the arm, the defensible position for a Chinese embodied company shifts away from the model layer.
The remaining moats are real-world data collection at scale, hardware reliability over thousands of hours, and deployment relationships with the factories and stores where the robot has to keep working. Those are slower to build and harder to replicate from a benchmark score.
The uncomfortable part for the sector is that the in-house foundation model, which was the fundraising story for two years, may become the least defensible layer of all.
Editor’s note: This is an adapted translation of the original 36Kr report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://36kr.com/p/4060707751826311.