PokeBot

PokeBot (Chinese: 破壳机器人, often shortened to 破壳; registered as 破壳时刻（上海）智能科技有限责任公司) is a Chinese embodied AI company founded in April 2026 by Tsinghua University assistant professor Huazhe Xu (许华哲), a co-founder of Galaxea Dynamics.^[4][6][7] The company works on robot manipulation and says its long-term goal is general-purpose robots for the home.^[1][6] On 30 October 2026 (Beijing time) it launched RoboAPI, a free, non-commercial technical-preview API that sends a robot's wrist-camera image, pose and a natural-language instruction to Poke-Aurora, the company's first world action model, and returns the next chunk of robot actions; the company says this works on compatible robots without task-specific data collection or training.^[1][2][3] PokeBot raised an angel round of tens of millions of US dollars in April 2026 and a pre-A round described as "亿美元级" (on the order of US$100 million) in August 2026.^[4][6]

Overview

History

Founding and angel round

Huazhe Xu is a tenure-track assistant professor at Tsinghua University's Institute for Interdisciplinary Information Sciences (IIIS), where he leads the Tsinghua Embodied AI Lab (TEA Lab). He received his PhD at Berkeley AI Research (BAIR) under Trevor Darrell and was a postdoctoral researcher at the Stanford Vision and Learning Lab under Jiajun Wu.^[20] Caixin described him as a co-founder of the humanoid robot company Galaxea (星海图), and AI科技评论 and OFweek said he had been Galaxea's chief scientist before founding PokeBot.^[4][8][10] In an interview with AI科技评论, Xu said PokeBot was founded by Tsinghua faculty and students who had previously worked together in his research group.^[8]

According to Caixin, PokeBot was founded in April 2026 and closed an angel round of tens of millions of US dollars (数千万美元) the same month. YunQi Partners (云启资本) led the round, with participation from Shunwei Capital, 弘晖基金, Xiaomi's strategic investment arm (小米战投), Galaxea, BV百度风投 (Baidu Ventures), 英诺天使基金, 水木清华校友种子基金 and 东方嘉富.^[4]

Cooking demonstration

In June 2026 PokeBot published a nine-minute video of a robot autonomously cooking mapo tofu, according to AI科技评论, which also reported demonstrations of shirt folding, threading cable ties and tying a sachet.^[7] The outlet described the company's technical approach as a "WAM x RL x DATA" flywheel: a world action model (WAM) for predicting physical change and planning actions, reinforcement learning on real robots, and real-world data gathered with self-built collection devices in home settings.^[7]

Pre-A round

On 3 August 2026 PokeBot announced a pre-A round that it called "亿美元级" (on the order of US$100 million), without giving an exact figure.^[6][7] Shunwei Capital and Matrix Partners (经纬创投) co-led it, with 九坤创投, 钧山资本, SEE Fund, 猎聘投资, 元诺资本 and 钟鼎资本 joining, and existing investors YunQi Partners, Xiaomi, 弘晖基金, 英诺天使基金 and 东方嘉富 adding to their stakes.^[4][6] Caixin's Chinese headline put the round at nearly US$100 million (近亿美元), while the headline of Caixin Global's English report said "hundreds of millions of dollars".^[4][5]

RoboAPI launch

PokeBot announced RoboAPI on X at 17:51 UTC on 29 October 2026, early on 30 October in Beijing ("Don't trust demos. Try it yourself."), and a GlobeNewswire release datelined Beijing and dated 29 October said the technical paper and product demonstration would follow on 30 October, Beijing time.^[2][3] The launch page, dated 30 October 2026, introduced RoboAPI and Poke-Aurora.^[1] The Python client package roboapi first appeared on PyPI on 27 October 2026 (version 0.4.2); version 0.4.5 was uploaded on 29 October 2026 and was the current release as of 1 November 2026.^[19]

Poke-Aurora

PokeBot describes Poke-Aurora as its first world action model, a model that learns to predict future visual observations together with robot actions so that patterns of object motion and interaction can transfer across objects and scenes.^[1] The company says that a world action model's ability to model the world "often does not naturally translate into reliable action generation", and that making video learning and action learning reinforce each other remains a key challenge.^[1]

Pretraining: Modulated Residual Coupling

For pretraining, PokeBot introduced a mechanism it calls Modulated Residual Coupling (MRC), which adds learnable interactions between the video and action modules while keeping each module's original computation path. The company says it observed that MRC improves gradient alignment between video prediction and action prediction, so that the model's world modeling contributes more to action learning.^[1] The launch page says more details of MRC will appear in a Poke-Aurora technical report; it does not give the model's parameter count or architecture details beyond this.^[1]

Post-training: a UMI-style configuration

To handle differences in arm structure, control method and sensor layout between robots, PokeBot took inspiration from UMI and post-trained Poke-Aurora on a single "UMI-style" configuration: a wrist-mounted fisheye action camera plus end-effector pose control.^[1] The company reports that switching camera brands (from an Insta360 Ace Pro 2 to a DJI Action 6) had almost no effect on performance, and that the model transferred to robot configurations it had not seen, such as moving from a UR7e to a YAM Standard arm. PokeBot concludes that a wrist-mounted action camera is enough for RoboAPI to control an arm, provided the arm has sufficiently capable inverse-kinematics solving.^[1] The launch page charts four configurations (UR7e/UR5e with a Robotiq 2F-85 gripper, Franka FR3/Panda with a Robotiq 2F-85, YAM Standard with a YAM gripper, and a floating-base Tianji with a YAM gripper) at 100%, 87%, 76% and 53% respectively, without stating what the percentages measure.^[1]

On training data, Xu told AI科技评论 that real-robot data makes up a small share, and that UMI-style and egocentric ("Ego") human data together account for 85 to 90 percent. He said most egocentric data is bought from outside suppliers, and that the company built its own UMI gloves and a small home-scene collection site in its Beijing office.^[8]

Evaluation

PokeBot built its own evaluation framework covering nearly 1,000 distinct instances of common manipulable objects and nearly 10,000 unique instructions, grouped into three task categories.^[1] It compared Poke-Aurora with "Astra" (identified in PokeBot's press release, and in a company-supplied release carried by 机器之心, as OpenAI's GPT-6 Astra), Physical Intelligence's π0.5 and MolmoAct2.^[1][3][9] All figures below are self-reported by PokeBot on a benchmark it designed and ran itself.

Source: PokeBot launch page, success rates.^[1]

The press release gave an overall task success rate of 74.2% and said this surpassed π0.5 "by 51.3%" and GPT-6 Astra "by 16.4%".^[3] PokeBot itself wrote that it "cannot guarantee" that users' experience with RoboAPI will match these statistics, and presented the public API as a way for outsiders to test the model's generalization.^[1] In the interview, Xu described the internal method as large-scale random testing in which about 10,000 randomly generated tasks were each run once, and said the random-test success rate was above 70% and pick-and-place success reached 80%.^[8]

RoboAPI

How it works

RoboAPI takes a robot's observations and state, follows a user instruction, and predicts the next action chunk in near real time.^[1] Developers can either run the RoboAPI client on a Linux computer connected to the robot, which handles observation, action requests and execution and offers a local browser console, or call the cloud model directly from their own code.^[11]

The cloud API uses a chat-completions interface that works with the OpenAI API Python SDK: the documentation's examples use that SDK, a production base URL of https://api.pokebot.com/v1, bearer-token API keys created in a developer console at platform.pokebot.com, and a POST /chat/completions request whose user message carries a JSON instruction and state plus a base64 camera image.^[12][13] Actions come back as a single execute_actions tool call whose output is an array of rows with eight values each: absolute flange-center position (x, y, z in meters), an xyzw quaternion and a normalized gripper opening.^[13][14] The RoboAPI local runtime dispatches each row as an absolute target at 30 Hz.^[14] Continuous tasks use sessions (POST /sessions), with the client returning the previous assistant message and an execution receipt (executed_steps, completed or interrupted) on each turn.^[14] The local runtime scales wrist-camera images from 1280x720 to 640x360 by default.^[14]

The client also exposes the robot to AI agents through the Model Context Protocol: version 0.4.5 provides 27 MCP tools, including observation, starting and stopping a policy, pausing and calibration.^[18] Developers can write adapters for other arms, parallel grippers and wrist cameras using a project generator in the package.^[23]

Supported hardware

The documentation warns that the camera's fisheye distortion is part of the model input, so other cameras must match the reference lens's projection and mounting.^[15]

Access, pricing and limits

During the technical preview, RoboAPI is free to everyone, but PokeBot describes it as "a non-commercial offering" that "does not provide products or services to for-profit third parties."^[1] The press release described it as a noncommercial offering "to eligible users" and linked to an early-access application.^[3] The developer console shows each key's model access, balance and usage.^[12] Xu told AI科技评论 that the company had sized its first-day deployment for at most about 100 concurrent users, that a robot API is hard to turn into a closed commercial loop because the installed base of robots is small, and that PokeBot chose a free API over open weights so that it could learn from real-world failures and keep improving the model. He also said he expects fewer embodied models to be open-sourced in future.^[8]

PokeBot's own usage notes say English instructions generally perform better, that occlusion from a single wrist camera can reduce performance, that RoboAPI keeps historical context and retries after failures, and that a safety gate rejects instructions that could harm people or surroundings. It recommends workspace safety boundaries or compliant arms and asks users to report dangerous behavior.^[1]

Testing before launch

PokeBot says it deployed robots in dozens of settings in its own facilities, and ran a private beta in which outside labs connected their own arms; the launch page shows external test videos labeled Xi'an, Xiong'an, Beijing, Shanghai, Shenzhen, Singapore, Princeton, Stanford and Boston.^[1] A company-supplied release carried by 机器之心 said the API had been tested in university labs in the United States, Singapore and elsewhere.^[9]

Home robot hardware

The company's website describes its robot hardware as "Coming soon".^[24] In the October 2026 interview, Xu said three prototypes came off the production line in September 2026 and that PokeBot planned to show them at CES the following year. He described the design as a wheeled humanoid with three-fingered dexterous hands (6 degrees of freedom), a lifting column that adjusts height from 1.2 to 1.7 meters, and a typical working height of about 1.4 meters.^[8] He said the company was not pursuing aggressive mass production, and suggested annual output from several hundred to a few thousand units as a sensible scale.^[8]

Comparison with other robot model APIs

RoboAPI launched in the same month as two other robot-model APIs. A PokeBot release carried by 机器之心 noted the trend without naming the companies.^[9]

Founder's views on general-purpose models

Asked about GPT-6 Astra, Xu told AI科技评论 that its strength showed PokeBot's focus on generality was right but made it a strong competitor. He said Astra was slow and expensive for robot control, could not yet do bimanual tasks such as folding clothes, and that smaller dedicated embodied models respond faster. He called himself part of a "resistance faction" (抵抗派) opposed to the view that large language models will absorb embodied models.^[8]

Continue reading

• WWorld action modelA world action model (WAM) is a robot policy design that builds action generation on a video world model backbone rather than on a vision-language model, so that a single network jointly predicts how a scene…Mentioned 6 times above
• UUniversal Manipulation InterfaceUniversal Manipulation Interface (UMI) is an open-source system for collecting robot manipulation training data with a handheld, camera-equipped gripper instead of an actual robot.Mentioned 5 times above
• PPhysical IntelligencePhysical Intelligence (also known as Pi or π) is an American artificial intelligence robotics company that builds general-purpose foundation models and learning algorithms designed to control any robot for any…Mentioned 3 times above
• Dream LabsDream Labs (legally Dream Labs, Inc.) is a San Francisco artificial intelligence company that builds robot foundation models and sells access to them through a public API.Listed in See also
• Galaxea DynamicsGalaxea Dynamics, also known as Galaxea AI (Chinese: 星海图, Xinghaitu), is a Chinese embodied AI and robotics company that builds wheeled dual-arm humanoid robots, standalone robotic arms, and the G0 family of…Listed in See also
• VVision-language-action modelA vision-language-action model (VLA model, or VLA) is a machine-learning model for robotics that uses visual observations and a natural-language instruction to generate actions a robot can execute.Listed in See also

More from See also

• Robot learning

More to read

• Tsinghua University
• Embodied AI
• Xiaomi (robotics)
• π0.5
• Mimic Robotics

Topics:Chinese AIEmbodied AIPhysical AIRobotics Companies

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22PokeBot. "RoboAPI: Zero-Shot Generalization, Now Within Reach". October 30, 2026. Chinese version: pokebot.com/zh-cn/ai.
2. ^1 ^2PokeBot (@PokeBot_Global). "Don't trust demos. Try it yourself. Introducing RoboAPI...". X, October 29, 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6PokeBot. "PokeBot Launches RoboAPI for Real-World Testing of Its Poke-Aurora World-Action Model". GlobeNewswire, October 29, 2026. Syndicated copy: The Manila Times.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Du Zhihang (杜知航). "破壳机器人创立4个月Pre-A轮融资近亿美元". Caixin (财新网), August 3, 2026.
5. ^Du Zhihang. "Chinese Robotics Startup PokeBot Raises Hundreds of Millions of Dollars". Caixin Global, August 5, 2026.
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9投资界. "破壳机器人完成亿美元级Pre-A轮融资". PEdaily (投资界), August 3, 2026.
7. ^1 ^2 ^3 ^4 ^5董子博. "独家 | 破壳机器人完成亿美元级融资：「操作」才是具身落地的最大挑战". AI科技评论 (via NetEase), August 3, 2026.
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10张含菁. "对话 | 许华哲：面对大模型的「入侵」，要做「抵抗派」". AI科技评论 (via NetEase), October 30, 2026.
9. ^1 ^2 ^3机器之心. "刚刚，破壳机器人发布RoboAPI！具身模型也能「开箱即用」了". 机器之心Pro (via NetEase), October 30, 2026. Company-supplied release (labelled 机器之心发布).
10. ^"一句指令，自己干活！「破壳机器人」发布RoboAPI！". OFweek 维科号 (author account 具身涌现), October 30, 2026.
11. ^PokeBot. "Get started". RoboAPI Documentation. Accessed November 1, 2026.
12. ^1 ^2PokeBot. "Account and API keys". RoboAPI Documentation. Accessed November 1, 2026.
13. ^1 ^2PokeBot. "Your first API call". RoboAPI Documentation. Accessed November 1, 2026.
14. ^1 ^2 ^3 ^4PokeBot. "Tasks and sessions". RoboAPI Documentation. Accessed November 1, 2026.
15. ^1 ^2 ^3 ^4 ^5 ^6PokeBot. "Hardware setup". RoboAPI Documentation. Accessed November 1, 2026.
16. ^1 ^2PokeBot. "Install RoboAPI". RoboAPI Documentation. Accessed November 1, 2026.
17. ^PokeBot. "Franka setup". RoboAPI Documentation. Accessed November 1, 2026.
18. ^PokeBot. "Connect an Agent over MCP". RoboAPI Documentation. Accessed November 1, 2026.
19. ^"roboapi". Python Package Index. Release history accessed November 1, 2026.
20. ^1 ^2Huazhe Xu. "Huazhe Xu (许华哲)". Personal homepage, Tsinghua University IIIS. Accessed November 1, 2026.
21. ^Dream Labs. "Zero-Shot to New Robots: Introducing Dream-1". October 2026.
22. ^Physical Intelligence. "Frontier Models for Every Robot: Announcing our API beta". October 13, 2026.
23. ^PokeBot. "Connect new hardware". RoboAPI Documentation. Accessed November 1, 2026.
24. ^PokeBot. "Hey! We're PokeBot.". Company homepage. Accessed November 1, 2026.

Show all 24 references