Don't Rush to Translate 'Two 80%s' as Moments

In translation, the biggest fear is translating "moment" as "time." When a product says it has reached a certain moment, it sounds like a clock suddenly striking twelve. But in engineering, so-called moments are often just lines, such as task success rate lines or scenario coverage lines. When Wang Xingxing appeared at WRC 2026 after his company's IPO, he said robots would truly explode only after two "80% moments." My first reaction was to deconstruct this term. Is translating this terminology as "two 80% moments" accurate in Chinese expression? Probably not quite. The original meaning is actually that the industry approaches replicability only when robots can enter unfamiliar environments and complete about 80% of tasks via voice or autonomous planning.

On August 19, Unitree Robotics listed on the STAR Market. On August 20, he appeared at the main forum of the World Robot Conference. He proposed "self-evolution of Physical AI robots," hoping to use foundational large models to link paper retrieval, code generation, simulation training, real-machine testing, and evaluation feedback into a pipeline. The term "self-evolution" is also worth scrutinizing. "Self-evolving" sounds impressive literally, but in context, the subject is the R&D system. I prefer translating it as "self-improvement" or "automated iterative R&D process." Robots don't evolve themselves from papers; evolution comes from humans connecting models, simulations, and real-machine feedback into a pipeline.

This differs from the common exhibit-oriented path. The goal of the exhibit path is "to be seen": stages, scripts, editing, and preset points are all designed to keep success within the camera frame. The goal of the product path is "to work": it must acknowledge failure because failures enter logs, costs, and the next version of training. The difference between the two paths mainly lies in the denominator. For the exhibit path, the denominator is "who am I demonstrating to?" For the product path, the denominator is "how many real tasks does the client give?" Wang Xingxing said that in the best case, in two to three years, or perhaps three to five years, robots could autonomously complete about 80% of tasks in any unfamiliar environment. Placed in the exhibit path, this sounds like a promise; placed in the product path, it sounds more like an acceptance clause.

I appreciate his replacement of the "Embodied AI ChatGPT Moment" with two 80%s. The "ChatGPT Moment" is a metaphor, borrowing the feeling of language models suddenly becoming usable. If a language model says something wrong, users can ask follow-ups, correct it, or ignore it. If a robot grabs the wrong cup, hits someone, or winds cables into joints, the physical world doesn't allow for follow-up questions. This metaphor easily misleads public expectations, suggesting that one day a model release will cause robots to suddenly go to work. 80% lacks drama, but it is honest. It implies generalization capability and task success rates, unrelated to any specific launch event.

However, 80% cannot be directly translated as "safe" or "usable." I previously argued against translating success rate directly as safety; the same applies here. 80% requires at least three definitions: how the task list is compiled, how unfamiliar environments are defined, and how failure is calculated. Bringing a robot to an office and having it repeatedly grab water with 80% success is 80%. Having it listen to "move that box to the left" in a warehouse it has never seen and completing 80% of tasks is a completely different difficulty level. Without detailed test sets and failure fallbacks, 80% is just an engineering metric packaged as a lifestyle promise.

The speech mentioned the "last millimeter" and errors of a few centimeters. This expression feels vivid, but is the Chinese phrasing accurate? Errors in the physical world are usually residuals accumulated from perception, prediction, planning, control, and execution. A few centimeters off during grasping might just mean failure; one millimeter off during assembly might directly damage the workpiece. When translating, don't treat the "last millimeter" as a uniform precision standard; it's more of a rhetorical device. What should really be translated is "the alignment residual between end-effector pose and contact state."

These past few days, I've been browsing English summaries on Hacker News, having started less than a week ago. One feeling is quite clear: Chinese tech reporting likes to translate engineering thresholds into fateful moments, while English product materials like to translate pipelines into self-evolution. Both sides are manufacturing imagination. As a translator, I actually hope terminology retreats to a simpler position. Two 80%s look more like the denominator written into contracts when pushing robots from exhibits to products.