Xingdong Era Closes Nearly 1 Billion RMB Series A+ Round; Stanford Team Unveils Home Robot Memo; π*0.6 Model Breakthrough | Radio Weekly

1. EraStellar Completes Nearly 1 Billion RMB in A+ Round of Financing

On November 20, EraStellar announced the completion of an A+ financing round totaling nearly 1 billion RMB. This round was led by Geely Capital, with strategic investment from BAIC Industrial Investment. The Beijing Artificial Intelligence Industry Investment Fund and the Beijing Robotics Industry Development Investment Fund (jointly managed by Jingguorui and Shoucheng Holdings) also provided joint capital injections; meanwhile, several international industrial giants have added their strategic capital support.

This financing will further support the technical iteration and commercial application of the end-to-end VLA embodied large model ERA-42. The involvement of Geely Capital and BAIC Industrial Investment also opens up new strategic synergistic development space for the company's industrial applications.

To date, the company has completed five rounds of financing. Prior to this current round, EraStellar announced the completion of an A-round financing of nearly 500 million RMB in July of this year.

In October 2024, EraStellar completed a Pre-A financing round of nearly 300 million RMB; in January 2024, it received angel investment exceeding 100 million RMB; and in October 2023, it completed a seed funding round amounting to tens of millions of RMB.

Founded in August 2023, EraStellar's embodied large model ERA-42 has achieved precise control over the full body and five-fingered dexterous hands of high-degree-of-freedom, full-size bipedal humanoid robots, with applications in logistics and commercial service sectors.

On the commercialization front, Xingdong Ji Yuan focuses on industry solutions for embodied AI in China. The largest single order in the logistics sector is nearly 50 million RMB, and standardized solutions have been established.

Abroad, the company targets the developer market, with overseas business accounting for 50% of its operations, covering North America, Europe, the Middle East, Japan, and South Korea.

2. Zhiyuan Expedition A2 Completes 100-Kilometer Cross-Province Walk

Last week, the Zhiyuan Expedition A2 successfully completed a cross-province walking challenge from Jinji Lake in Suzhou to the Bund in Shanghai. Its quick hot-swappable battery system allowed the Expedition A2 to remain powered on and continuously operational throughout the journey.

On November 20, the Guinness World Records adjudicator presented a certificate to Zhiyuan Expedition A2, officially confirming that Zhiyuan Expedition A2's total walking distance was 106.286 km, making it the Guinness World Record holder for "the longest distance walked by a humanoid robot."

Regarding issues such as autonomy and stability during this challenge, Wang Chuang, Partner at Zhiyuan Robotics, Senior Vice President, and President of the General Business Division, answered questions from media outlets including 42nd Radio Wave (the following content has been edited without altering the original meaning).

The A2 used for the challenge was taken directly from the factory floor; it underwent no special modifications and is identical to the product available to users.

In terms of challenge format, A2 adopts a hybrid approach combining human assistance with some autonomous robot functions, which can be understood as semi-autonomous. Zhiyuan hopes that this challenge will draw the attention of the entire industry supply chain to issues regarding the stability and reliability of robots operating over extended periods.

3. Embodied AI Startup Sunday Launches Home Robot Memo

On November 20, Sunday, an embodied AI startup co-founded by Stanford University PhDs Zhao Zihao and Chi Cheng, launched its first wheeled humanoid robot, Memo.

Designed primarily for home use, Memo can fold socks, clear tables, and brew coffee. The company claims it can acquire skills faster than any previous robot, with a beta version scheduled for release by the end of 2026.

Memo features a minimalist, soft, anthropomorphic design. It stands 1.7 meters tall, weighs 77 kilograms, and has a shell made of silicone material with rounded corners.

Its proprietary Skill Capture gloves are embedded with high-precision sensors capable of capturing three-dimensional data such as finger joint angles, force exertion, and motion trajectories. When "memory developers" wear the gloves to perform household chores, the data is uploaded in real-time to Sunday's cloud-based AI model, processed, and converted into mechanical commands executable by Memo.

Sunday plans to provide Memo robots to 50 households by the end of 2026. These early adopters will gain early access and work alongside the team to guide Memo in learning new skills. Before product delivery, Sunday will utilize beta testing to ensure the robots meet the highest standards.

4. π*0.6 Model Enables Breakthrough in Robot VLA

On November 18, Physical Intelligence (hereinafter referred to as PI) released its latest foundational model, π*0.6. Through an innovative Recap training method that integrates demonstration, guidance, and autonomous practice, the Visual-Language-Action (VLA) model has broken through the bottleneck of imitation learning.

In terms of specific operational mechanisms, the key to π*0.6 is the Recap training method, which allows robots to go through three stages: 'demonstration learning, error-correction guidance, and autonomous practice.'

The model can extract effective training signals from 'bad data,' rather than simply having robots repeat certain actions through demonstration learning.

In the error-correction guidance stage, remote operators take over when the robot makes a mistake, demonstrating how to recover from errors. This targeted error-correction data directly addresses special problems the robot encounters in real-world scenarios, breaking the cycle of error chain reactions.

Additionally, Recap introduces value functions to solve the 'credit assignment' problem in reinforcement learning, i.e., determining which actions led to final success or failure. By predicting task completion probabilities across different scenarios, the model identifies critical effective actions and error nodes, thereby reinforcing advantageous behaviors and avoiding erroneous operations, enabling continuous improvement through autonomous practice.