Embodied AI and Humanoid Mass Commercialization
Evaluating generalization capabilities, training datasets, and mass manufacturing paths for humanoid robots.
3 sourcesVoices in this topic
All paragraphs
Oct 26, 2026
Oct 24, 2026
- 王淏楠 淏楠 favors an embodied-AI split between a large-model brain and proprietary hardware with smaller models. He says embodied-AI companies have “two orders of magnitude” fewer GPUs than large-model companies. A more realistic startup moat is to …
- Astra’s most important potential value is taking over failed steps and closing the data loop. 淏楠 imagines that once a small model reaches a 50%–60% success rate in unfamiliar settings, GPT can take over failures and generate …
- Embodied AI’s high valuations still need to be earned in real-world settings. Asked by 曲凯 about a peak valuation of around $20B and roughly RMB5B–RMB6B on the balance sheet, 淏楠 pointed to hotel bed-making: flat floors …
Oct 21, 2026
- 钱程 Qian sees long-term potential in embodied robotics, but rejects the idea that today’s technology buzz translates directly into near-term household revenue. He estimates that some commercial and factory applications could be ready in 3–5 years, while consumer-home applications may take 5–10 …
- ECOVACS envisions a future with multiple specialized types of household robots, not one in which humanoids swallow every tool category. Qian asks whether consumers, once robot vacuums are widespread, will really “buy another vacuum cleaner for a robot.” The company has not …
Oct 19, 2026
Oct 17, 2026
- Emad Mostaque The cost for a frontier AI model to complete a physical robot subtask fell 84% in two months — from $2.36 in July to $0.38 in September on YC startup RoboCurve’s RoboDojo RC Tier 1 benchmark. The panel discusses task costs halving every three weeks; Peter Diamandis says that trend could make robotic intelligence effectively cost …
- Robots’ hardware and supply chain may become the constraint as model costs fall. Salim Ismail says entrepreneurs should build robots, actuators, and spare parts. Figure’s Brett Adcock has committed $1 billion for data, …
Oct 15, 2026
- Lindon Gao Dyna founder Lindon Gao calls the company a “Product Research” outfit and describes its approach as “all-out.” He says product loops should steer research: deploying a task is not the same as closing a commercial loop, and “robots needing a babysitter …
- Dyna is betting on egocentric, first-person data, and says it has found a “very strong scaling law” in a pretraining dataset at the 1-million-hour scale. The company has shifted from prioritizing high-quality data to seeking broader, more diverse data. Lindon admits that he selectively …
- His contrarian commercial view is that demand for general-purpose B2B robots will exceed demand for B2C home robots; the home is a difficult, distant product direction, not the first market to tackle. His case: a 4.5% US unemployment rate means businesses employ 95.5% of the US workforce; assuming a household of 3, the US may need only …
- His view on “brains vs. bodies”: the brain sets the floor for what a robot can do; the body sets the ceiling for how well it can do it. He believes supply chains may matter more over a 10-year horizon. Dyna has a 40–50-person hardware R&D and supply-chain team in China, and …
- He estimates Dyna is still 1–2 years from product value that can scale, and expects the industry to converge further over the next 1–2 years. He points to GPT-3, released in API form in June 2020, and ChatGPT, launched as a chat product in November 2022, to show why product form …
Oct 14, 2026
- 陈哲 The quarter’s biggest shock was GPT-6 Astra, which showed that general models can enter physical AI: frontier labs are inclined to solve AGI first and treat physical AI as a subproblem. Astra does more than break tasks down; it can directly output end-effector poses, which are converted into motor signals through inverse …
- Chen Zhe sees Astra’s impact in 3 layers: directly generating policies, supervising and correcting VLA, and—most “hair-raising” of all—acting as a robotics researcher. In Galaxy General’s case, Astra directly controlled 10 Robot Dojo tasks with a 26% success rate; supervising π0.5’s outputs raised that to …
- Chen expects every leading general-model company to make moves in robotics: there are rumors around Kimi and Zhipu, Liang Wenfeng has talked about physical AGI, and teams such as ByteDance Seed and Tencent RoboticsX may also pivot toward a generative-AI approach. The compute gap remains vast: leading embodied-AI companies may have GPUs in the thousands, versus hundreds of thousands or even millions at …
- The embodied ICL capabilities shown by GEN-1.5 and S1 could cut the post-training needed for deployment to minutes of demonstration video. Two surprising findings: out-of-distribution tasks may be learned from a single demonstration, and long sequences of combined skills—such as …
Oct 13, 2026
- Jason Ma Dyna Robotics claims a proof point many robotics startups lack: validated multi-site commercial scale, not just pilots. Its robots fold thousands of napkins per day across many Din Tai Fung locations, and Dyna also folds laundry for Best Western. Jason says …
- Vibhu Task adaptation takes tens of minutes to a couple of hours of robot data collection. The chopping task took 45 minutes; opening the fridge is an engineer-onboarding interview task. Vibhu says the new robot learned tasks …
- Dyna 2.1 is presented as hardware with a “physical agentic harness” for longer workflows. Vibhu says a complete laundry loop can take about an hour, and argues that robotics’ value lies in workflows that run unsupervised for …
- Dyna’s approach is full-stack: it builds its own inference stack and is working toward owning its actuators, alongside its semi-humanoid hardware with a wheeled base and 4-DOF waist. Vibhu says legs bring trade-offs in energy use, stability, and safety, and argues that iterating a semi-humanoid with real customers is …
- Jason Ma Deployment targets are tens of robots already and hundreds next year, with millions of robots as the longer-term vision. Jason warns that “there’s a whole graveyard of robotics companies” with shiny demos, and says teams focused on only part of the problem risk …
Oct 7, 2026
Oct 3, 2026
- Keerthana Gopalakrishnan Keerthana reads China’s viral Robot Olympics as genuine locomotion progress but a poor proxy for near-term robot utility. Running on flat, rigid terrain is unusually simulation-friendly, while useful manipulation breaks down around contact, friction, deformable …
- Robotics remains at its “GPT-2 moment” because impressive demonstrations have not yet become portable, reliable intelligence. GPT-3-level progress would require learning from multiple examples across many tasks and a brain that transfers across humanoids, arms, and …
- Gemini Robotics 2 separates embodied reasoning from physical execution across three models. Gemini Robotics ER2 is the higher-level, tool-using “brain”; Gemini Robotics 2 is the vision-language-action model controlling motion; and …
- Controlled commercial environments remain the likeliest early deployment wedge, even though model latency may matter less at home. Nathan emphasizes repeatable factory and warehouse tasks and professional servicing; Keerthana agrees that commercial use cases are more …
- Cross-embodiment is improving rapidly, but “one brain, any body” is still an aspiration rather than a solved capability. Google DeepMind’s models now control humanoids “from fingertips to feet,” and its on-device program reportedly adapts tasks to new bodies …
- Hardware has moved faster than Keerthana expected, shifting the remaining work toward reliability, repeatability, cost, and safe force control, while dexterous hands approach a capability bottleneck. In roughly a year and a quarter, the frontier advanced from grippers to multifinger control, garbage-bag tying, and whole-body manipulation. …
Oct 1, 2026
Sep 29, 2026
- 熊友军 熊友军 defines humanoid robot track meets as an extreme testing ground: the competitions are not a substitute for industrial deployment, but a way to find the limits of the body, power system, motion control and reliability against a common yardstick. Tiangong Omni improved from 9.44 seconds in rehearsal to 8.64 seconds, as the density of competition kept pushing both software and hardware …
- The viral “faceless run” was not a preprogrammed performance, but a solution the robot found on its own while chasing maximum speed under shoulder-joint and power constraints. The team initially disliked the posture as “not that pretty” and even considered retraining it, but kept the fastest version: the robot gave …
- The most important annual change at the track meet was not speed but autonomy: last year, Tiangong was the only one of more than 200 robots to compete fully autonomously; this year, track-and-field robots had largely left the remote controls behind except in some obstacle events. 熊友军 sees this only as “a demonstration of progress in the brain,” because recognizing lane markings, taking bends and adjusting running form …
- Embodied intelligence may ultimately require data on the scale of tens of millions of hours, but 熊友军 refuses to treat “throwing more at it” as the inevitable end state. Today’s general-purpose motion controllers use only hundreds of hours of human data, while world-model training runs to roughly tens or …
- The real threshold for training a robot’s “brain” is a combination of compute, data and learning paradigm—not the launch of another model name. 熊友军 agrees that 1,000-GPU effective compute is a meaningful benchmark; with Baidu and multiple cloud platforms, the innovation center …
- Near-term commercialization will still begin in structured industrial settings, but those settings are better understood as a bridge to household robots than as the final destination. In parcel sorting, the fastest humans handle more than 2,000 items an hour, FANUC was cited at more than 1,500, and domestic robots at 1,000 …
Sep 20, 2026
- 徐梦迪 徐梦迪 says he is seeing early signals of the Scaling Law in embodied intelligence, but stresses that loss does not equal success rate. He cites results from other companies showing that scaling pretraining data from 100K hours to 1M hours produced strong performance on …
- He sees embodied intelligence as being at the GPT-1 stage today—pretraining plus fine-tuning for a target domain—with the key inflection point being a GPT-3-style in-context learning moment. The end state is for end users to define robot behavior with prompts and demos: clothes could be folded into squares, rolled up, or simply …
- His overall view of China’s embodied-intelligence startups: “There has been a great deal of real progress, but there will also be some bubbles; the two are not mutually exclusive.” Chinese companies are working across the full stack—materials, sensors, robot bodies, data, and models—with resources continuing to flow …
- The riskiest part of commercializing embodied intelligence is the chicken-and-egg problem in the data flywheel. The base model has to be good enough for an end user to see “obvious progress” after 30 minutes of teaching before the loop can start …
- On data, he calls himself a “fusionist” and sees very high value in human data. Teleoperation data is closest to the robot’s embodiment and suited to post-training; simulation data is clean and controllable and suited to …
- The conversation was threaded through with methodological one-liners. 李飞飞 first pressed him on which benchmark metric was “actually important”; 徐梦迪’s answer was final success rate. She also quoted Einstein: …