ACE Robotics Fills the Data Gap, Then Hits Cost

ACE Robotics says 1,000 workers in wearable kits can log 10,000 hours of robot-training data in a day. The SenseTime-backed firm, chaired by co-founder Wang Xiaogang, argues that data scale, not another arm or camera, is what still separates useful machines from demos.

The same company also says the combined cost of hardware, compute and deployment is the hardest unsolved problem in embodied AI, and Wang has now dated broad commercial use years past a late 2027 inflection.

Wearable Kits Replace Hours of Robot Teleoperation

ACE, founded in July 2025, unveiled Ambient Capture Engine 2.0 on 19 July 2026 at a forum it ran during Shanghai’s World Artificial Intelligence Conference. The pitch is blunt. After years of work, the industry has about 100,000 hours of manipulation data, most of it from teleoperation, in which a person drives a robot while the system records the motion. That method ties up costly hardware and trained operators for every hour in the log.

The new capture stack is built around people already doing the work. The ACE Ego Kit is a wireless wearable with head, hand and chest units, including the ACE Sense Glove. The glove is rated to 0.01 newtons of force sensitivity and less than two degrees of joint-angle error, and the kit keeps more than 20 mixed sensors inside one millisecond of one another.

EGO KIT CAPTURE SPECS

• Force floor: The Sense Glove is specified down to 0.01 newtons of tactile sensitivity.
• Joint error: Hand tracking is listed at less than two degrees of angle error.
• Sensor sync: More than 20 mixed sensors stay inside a one-millisecond timing window.
• Throughput claim: 1,000 people in the kits, the company said, can produce 10,000 hours of real-world data in a day.

Annotation sits in a companion ACE Data Engine aimed at long-horizon tasks, with ACE Ego Matrix meant to line up data across operators, devices and robot bodies. The company wants to move from those 100,000 hours to the tens of millions it says the field needs. At the stated 10,000 hours a day, 10 million hours would take 1,000 days of that pace, and “tens of millions” would take longer still. The kits do not retire the humans. They put the humans in sensors so the robots do not have to be the cameras.

The 8-Billion-Parameter Brain on Jetson Thor

Kairos 3.1 is the world model ACE showed with the capture kit. A world model gives a robot an internal picture of a scene so it can plan a move. The 19 July statement describes a hybrid Transformer that folds vision, language, force, touch and motion policies into one latent space, then runs an understand, reason, execute and reflect loop. In one internal test, the robot switched from a three-finger grasp to a four-finger grasp after a miss and finished the pick. Spotting the miss is the easy frame. Issuing a safe corrective motion, at full speed beside people, is the step most demos still skip.

In the digital world, a model error may result in a flawed image or paragraph. In the physical world, an incorrect action can have real consequences.

Wang Xiaogang, Chairman, ACE Robotics, WAIC 2026 forum

In internal tests, the Kairos 3.1 8 billion parameter edition hit 125 milliseconds of inference on Nvidia’s Jetson Thor at BF16 precision, using ACE’s KairosRT engine on the robot. The company told interviewers the robot does not need a live link to perceive, decide or act safely, with the cloud kept for updates, fleet coordination and rare cases. That 8 billion parameter build is not the same model ACE open-sourced in March 2026. Kairos 3.0-4B is a 4 billion parameter world model the firm said ran 72 times faster than Nvidia Cosmos 2.5 on an A800, used 23.5GB of VRAM against Cosmos 2.5’s 70.2GB, and generated at a 1:1.5 ratio of compute time to video on a Jetson Thor T5000 listed at 517 TFLOPs.

Nvidia is on the same on-device path. Its on-device Cosmos 3 Edge model is also a 4 billion parameter world model aimed at Jetson Thor, so a small brain on the machine is no longer a private trick. ACE says Kairos also runs on other processors, including boards from Rhino Tech and Digua Robotics, so a warehouse buyer is not locked to one chip vendor. Spatial understanding is handled by ACE-BRAIN-0.5, which the firm said led 12 public tests as of July 2026. A related generator, Kairos-HomeWorld, is built on 300,000 residential floor plans, 5,000 simulated homes and 8,700 3D assets tuned to common Chinese layouts, a geographic bias that will show up if the same weights are dropped into a different housing stock.

150 Public Hours Against a 10,000-Hour Pitch

The capture marketing and the public corpus are different objects. With Nanyang Technological University, ACE released ACE-Data-0 at the end of July 2026, a 150-hour household interaction set with 17 million video frames, 200 task categories, 50 participants, two homes and 75,000 episodes. Ego and multi-view video, body and hand motion, object trajectories, audio and touch sit on one timeline. That is dense, aligned data. It is also 150 hours, not 10,000 in a day.

THE HOURS ON THE TABLE

Open X-Embodiment remains the pretraining pile for many vision-language-action models, at more than a million real-robot episodes across 22 bodies. DROID, the cleanest single-arm public set, is 350 hours of Franka teleoperation over 76,000 trajectories. Neither is measured in the same unit as ACE’s 100,000-hour industry figure, and neither looks like 10,000 hours from one shift of wired-up staff. The wearable claim is a production rate. The public release is a research slice. Closing that gap is the whole product, and it is still a claim about capacity, not a dump of tens of millions of hours.

Xiaoman Runs in Live Convenience Stores

The commercial layer is three named lines, not a general-purpose humanoid sold as a box. Xiaoman is instant-retail fulfillment, paired with ACE’s W1 robot. Xiaoxin is hotel laundry. Xiaotu is outdoor work on quadruped bodies under a “one brain, many bodies” setup. Wang has also said the software is being put on humanoids from Unitree, AgiBot and Fourier in unmanned stores, hotels and instant-delivery warehouses.

THREE LINES IN THE FIELD

• Xiaoman: Instant retail fulfillment on the W1, which needs a 75 centimeter aisle, holds a robot-to-payload weight ratio under 2:1, and force-controls within one newton on both rigid goods and soft packs.
• Xiaoxin: Hotel laundry covering collection, wash, sort and fold, aimed at the night window ACE puts at about 80% of demand, with ironing listed as a later add.
• Xiaotu: Outdoor guidance and patrol on quadruped platforms, already used at the Begonia Flower Festival in Tianjin, with charging, dispatch and multi-robot coordination in the spec.

Xiaoman is the line with named buyers. ACE said it is live with Sense MartGo, Kuaikeda and PetroChina convenience stores, using real orders and shelf maps, and that it has also been run in an Alibaba Taobao Instant Commerce warehouse in Hangzhou. SenseTime Shanhui, PetroChina and Kuaikeda are listed as retail partners for autonomous warehouse-stores and unmanned shops. The expansion math is a plan: 1,000 retail locations within a year of the July launch, and about 10,000 stores within two years. Those figures are targets. They are not a current site count.

W1 is sized for dense shop aisles, not a wide warehouse boulevard, which is why a 75 centimeter minimum matters. Deployment is described as a matter of days into flash warehouses, small stores and warehouse-store hybrids. Cloud partners named on 19 July were Baidu AI Cloud, Alibaba Cloud, Huawei Cloud, Tencent Cloud and SenseCore AI Cloud. Compute diversity is part of the cost story, not a footnote.

Cost Caps Scale After the Data Rush

ACE does not pretend the kits finish the job. For mainstream sites, it says the gap between simulation and reality has narrowed a lot. What remains is “the extreme long tail of open, unstructured environments.” Then comes money.

For embodied AI to reach genuine commercial scale, the combined cost of hardware, compute, and deployment operations must fall below the threshold the industry can absorb. It is as much an industrial problem as a technical one.

ACE Robotics, company statement

Omdia expects the embodied AI market in Asia-Pacific to reach $1.5 billion in 2026. Chief analyst Lian Jye Su, in the firm’s general-purpose embodied robot study, still puts the field in a hype phase, with large rollouts weighted to consumer gadgets rather than enterprise fleets. A 10,000-hour capture day does not change a robot’s bill of materials, the night crew that still has to reset a jammed fold, or the integration work inside a PetroChina shop that was not built for a 75 centimeter aisle robot.

Money around the startup is real and still early. SenseTime’s newsroom said ACE closed an angel round led by Ant Group on 10 February 2026, with Qiming Venture Partners, JinJing Capital, Hony Capital, Lenovo Capital, Shanghai Jiao Tong University’s Hanyuan Asset and SenseTime Guoxiang Capital in the list. Wang said the firm raised more than $100 million in the first half of 2026 across several rounds and wants a listing “as early as permitted,” against Chinese rules that usually want three fiscal years of operations. A July 2025 founding date makes that listing a later chapter.

Wang Dates Broad Use After 2027

In late August 2026, Wang said he expects a “ChatGPT moment” for embodied intelligence by the end of 2027, driven by world models and environmental capture. He added the qualifier that matters more than the slogan. Even if that inflection arrives on time, he said, it will likely take another four to five years to see broad commercial implementation of embodied world models across sectors. That is the second clock. Data collection can be industrialized with gloves and headsets. Store-by-store cost, safety and the long tail still sit on a calendar that runs into the next decade.

THE BUILD-OUT SINCE 2025

1. July 2025: ACE Robotics is founded, with Wang Xiaogang as chairman and backing from SenseTime.
2. 18 December 2025: The venture steps out with the open Kairos 3.0 world model and a human-centric capture pitch.
3. 10 February 2026: SenseTime’s newsroom posts the Ant-led angel round.
4. March 2026: Kairos 3.0-4B is open-sourced as a 4 billion parameter on-device world model.
5. 19 July 2026: WAIC forum launches Kairos 3.1, Ambient Capture Engine 2.0, Xiaoman, Xiaoxin, Xiaotu and the PHYSICAL IQ benchmark with more than 20 university and industry partners.
6. 30 July 2026: ACE-Data-0 goes up as a 150-hour public household set.
7. 1 September 2026: With the University of Hong Kong, ACE posts StreamPI, a temporal add-on for vision-language-action models that does not add parameters.

StreamPI is the software move after the data launch, aimed at continuous temporal memory rather than another hour of footage. ACE published it on its own account.

ACE Robotics, in collaboration with The University of Hong Kong @HKUniversity, introduces StreamPI.

StreamPI brings continuous temporal understanding to Vision-Language-Action models, enabling robots to move beyond single-frame decisions and understand evolving physical worlds:… pic.twitter.com/E2WezwA0l2

— ACE ROBOTICS (@ace_robotics) September 1, 2026

A thousand wired-up workers can, on ACE’s numbers, outrun a year of robot teleoperation in a single shift. The 8 billion parameter brain can, on ACE’s tests, answer in 125 milliseconds without a datacentre in the loop. Wang has already said what follows that trick: an inflection he puts at the end of 2027, then four to five years of industrial grind before embodied world models are a broad commercial fact.