> Published 2026-07-08/10 · Unitree Robotics + UCSD · Embodied AI × Medicine > Original report: https://new.qq.com/rain/a/20260709A097ZO00 > Nature paper: https://www.nature.com/articles/s41586-026-XXXXX (published online July 8) > Secondary coverage: https://tech.ifeng.com/c/8uc57Rhz6ZR | https://new.qq.com/rain/a/20260710A05SHA00
The Event: Not a Demo, but a Nature Paper
On July 8, *Nature* published online a paper from the lab of Professor Michael Yip at UC San Diego: *"Feasibility Study of Humanoid Robots for In-Vivo Surgery."* The team used two Unitree G1 humanoid robots (nicknamed Surgie) to perform complete laparoscopic cholecystectomies on two live pigs.
Several engineering details stand out:
1. Standard commercial instruments, not custom robot tools. A connector mounted on the G1's wrist allowed it to hold wristed laparoscopic instruments from Korea's LivsMed—the same instruments used by human surgeons. Opening/closing angles were mapped to 0–45 degrees via a servo interface.
2. Two different collaboration modes across the two procedures. The first surgery was "one robot + bedside human assistant"; the second was "two G1 units standing side by side, purely robotic collaboration." Neither procedure converted mid-surgery to conventional laparoscopic or open surgery—a hard clinical criterion for "genuinely robot-performed."
3. The entire teleoperation framework was designed by a Chinese PhD student. First and corresponding author Lucas Zekai Liang graduated from Huazhong University of Science and Technology (Qiming College) in 2023 and is now a PhD student at UCSD under Professor Michael Yip (ECE department). The second surgery took 32 minutes.
4. Unitree had no prior knowledge of the project. According to an interview by Zhejiang Business Magazine with a Unitree representative, the two G1 units were purchased independently by the overseas research team, with no communication with Unitree beforehand. This means it is not a "vendor PR showcase" but an independent academic stress test of a consumer-grade humanoid robot.
Why It Matters: Three Fronts Advanced at Once
1. Embodied AI touches the "clinical usability" threshold for the first time
Embodied AI has been repeatedly validated in industrial, home, and special-operations settings—but clinical surgery has always been the ceiling-level scenario. Not because of technical difficulty, but because of a different error tolerance: a misgrasp in a factory can be retried; a 1 mm deviation on the operating table is a medical incident.
The paper demonstrates that general-purpose humanoid robots (consumer-grade hardware) + standard laparoscopic instruments (hospital stock)—two "non-specialized" conditions combined—can still complete a clinically standard full procedure on a live animal. It is an engineering-level "compositional validity" claim: not "Unitree robots can do surgery," but that the combination of consumer humanoid + standard instruments + skilled surgeon teleoperation is feasible.
Shanglei Liu, UCSD assistant professor of surgery and co-corresponding author on the second surgery, put it bluntly: "The cost is a fraction of the da Vinci's, and so is the footprint. Easy to deploy from remote rural areas to battlefields, even space."
A concrete comparison: the G1 is about 1.52 m tall, ~27 kg, and the whole system may cost only ~5% of a da Vinci system. Da Vinci (Intuitive Surgical) monopolizes the surgical robot market with 10,000+ installed units starting at $2 million each. If this path works, developing countries, field hospitals, space stations, and hospital ships gain a lightweight surgical robotics option for the first time.
2. It closes the loop with Unitree's IPO registration on 07-03/04
Seen alone, this news is easy to underestimate. Placed next to the past 10 days of Unitree news, the pattern is clear:
- 07-03: Unitree's STAR Market IPO prospectus discloses fully in-house joint actuators at one-third the cost of imports
- 07-04: CSRC approves Unitree's IPO registration
- 07-08/10: Nature paper—third-party academics perform in-vivo surgery with the G1
- Multiple Chinese hospitals and US academic institutions launching similar preclinical studies
- Embodied-AI companies investing heavily in fine manipulation, medical compliance, and low-latency teleoperation
- Surgical robotics shifting from dedicated devices (da Vinci model) to "general platform + medical adaptation layer" (G1 + medical kit model)
In a single week, Unitree obtained a public-market ticket, third-party academic validation, and clinical feasibility evidence—the first time a Chinese embodied-AI company has been certified on all three dimensions in the same week. The triangle of "public benchmark + academic endorsement + clinical proof" is taking concrete form.
3. The engineering baseline for "embodied AI × medicine" is raised
The paper notes limitations: the G1 still requires repeated corrections and does not yet meet surgical sterility standards. But these are engineering-solvable problems, not fundamental physical or algorithmic barriers. Over the next 12–18 months we will likely see:
Risks and Open Questions
Some honest caveats:
1. Live pigs ≠ human clinical trials. Porcine anatomy and response to laparoscopy differ from humans, and the paper explicitly requires repeated corrections. This is "surgery completed under skilled surgeon teleoperation," not "independent robotic surgery." Calling it autonomous surgery is over-interpretation.
2. The core contribution is engineering feasibility, not an algorithmic breakthrough. The system uses a classic master-slave teleoperation framework, not AI autonomous decision-making. The paper does not quantify AI's share of surgical decisions—entirely different from paradigms like NVIDIA's ASPIRE (AI-generated code driving actuators).
3. Unitree's stock volatility. The Nature paper is positive news, but translating it into clinical application takes 12–18 months. Markets may overreact short-term and be disappointed later if medical compliance moves slower than expected.
4. Intuitive Surgical's response. Its moat is not hardware but 20 years of clinical data, surgeon training systems, and hospital compliance pathways. A general platform like the G1 needs FDA/NMPA approval to truly threaten it—a timeline of at least 3–5 years.
Overall, this is a landmark story of embodied AI stepping from the lab into a real high-value scenario—but real commercialization is still years away. For today's market sentiment, it is a shot in the arm for the embodied-AI sector; for the engineering reality six months from now, it is a direction worth tracking long-term.