ACE Robotics Turns Workers Into Data Engines for Embodied AI

ACE Robotics says a thousand people wearing its camera headsets, tactile gloves and motion suits can generate 10,000 hours of real-world manipulation data in a single day. That rate aims to close an industry shortfall the company puts at roughly 100,000 hours of usable footage today, far short of the tens of millions needed for reliable embodied AI.

The Ambient Capture Engine 2.0 system, shown at the World Artificial Intelligence Conference in Shanghai in July 2026, records ordinary workers in kitchens, warehouses and hotels instead of forcing operators to teleoperate expensive robots. The shift removes robots from the collection loop and feeds the company’s Kairos 3.1 on-device world model.

By treating everyday labor as the capture surface, the company argues it can raise both volume and information density at once. The same workers who stock shelves or fold laundry become the source of the action-state pairs that imitation learning and on-robot world models require.

The Binding Constraint Was Never Just Hardware

Wang Xiaogang, SenseTime co-founder and ACE Robotics chairman, frames the problem bluntly. “The binding constraint for the humanoid and embodied AI industry today is data scale,” the company told Computer Weekly. Years of effort produced only about 100,000 hours of manipulation data, much of it from teleoperation that ties up costly robots and skilled operators for every recorded hour.

Teleoperation remains the gold standard for precise action-state pairs, yet throughput is low. Skilled operators often manage only 5 to 50 episodes per hour before fatigue degrades quality. Existing public datasets show the fragmentation: many capture video without force or tactile signals, or they stay limited to lab tables and short clips rather than long-horizon household or warehouse sequences.

ACE Robotics classifies data by information density from L1 to L5. Higher levels add three-dimensional force, tactile signals, failure-recovery trajectories and open-environment variables. Only that L5 material, the company argues, supports generalization, reflection and continuous learning.

• Industry baseline: ~100,000 hours of manipulation data accumulated after years of effort
• ACE daily claim: 1,000 wearers produce 10,000 hours of real-world footage
• Target scale: tens of millions of hours for intelligence to emerge at volume
• ACE-Data-0 release: 150 hours, 17 million frames, 75,000 episodes across 200 task categories

The gap between today’s stockpile and the stated target is not a modest shortfall. At the claimed wearable rate, a single day of thousand-person capture equals a tenth of everything the industry has gathered so far. Sustaining that pace is how the company positions ambient recording against years of robot-bound teleoperation.

The ACE-Data-0 paper with 150 hours of home tasks describes how the system turns real homes into calibrated recording studios at table scale for fine hand-object work and room scale for whole-body activity. Fifty participants performed goal-level instructions rather than rigid scripts, preserving natural hesitation and planning.

That choice matters for long-horizon learning. Scripted demos rarely include the pauses, re-grips and partial failures that later appear on a live floor. Goal-level prompts keep those behaviors in the record so the downstream model can train on recovery as well as success.

Wearables That Record Force Down to 0.01 Newtons

Ambient Capture Engine 2.0 centers on the ACE Ego Kit, a lightweight wireless set of head, hand and chest components. The ACE Sense Glove claims force sensitivity to 0.01 newtons and joint-angle error under two degrees. More than 20 heterogeneous sensors stay synchronized to under one millisecond.

Sub-millisecond sync is what keeps multi-modal streams usable as training pairs. When force, joint angles, egocentric video and body motion drift apart, imitation learners inherit timing noise that later shows up as unstable grasps. The kit’s design goal is to hold that drift inside a window tight enough for on-device world models to treat the stream as a single coherent episode.

Supporting software includes ACE-ViDiHand, a generative 4D hand-motion framework that the company says led three public benchmarks with 0.997 frame-level accuracy and up to 4.8-fold smoother motion. ACE Ego Matrix standardizes data across spatial coordinates, embodiment structure, action timing and quality so different operators and platforms can reuse the same streams. The framework has been open-sourced.

An automated ACE Data Engine handles continuous high-precision annotation of long-horizon tasks. The company also released ACE-Data-0 as an L5 household-interaction set. Full technical notes and access appear on the project page and dataset release notes.

1. ACE Ego Kit: headset camera, Sense Glove, chest unit for egocentric and body signals
2. Multi-view capture: synchronized egocentric plus exocentric video, full-body and hand mocap, object 6-DoF trajectories, audio and tactile
3. Annotation layer: automatic derivation of poses, meshes, bounding boxes and event descriptions from tracked states
4. Standardization: ACE Ego Matrix for cross-platform reuse and leaderboard placement on RoboCase and RoboTwin

The approach deliberately leaves robots out of data collection. A person simply works; the kit records the full perception-action loop that imitation learning and world models need.

Open-sourcing the matrix layer is meant to lower the cost of reuse across embodiments. Without a shared coordinate and timing scheme, each new robot arm or gripper forces another round of manual alignment. Standardized streams let the same household episode feed multiple platforms and public leaderboards.

Kairos 3.1 Runs the Loop On the Robot Itself

Kairos 3.1 is presented as a unified action-oriented world model rather than a stitched stack of separate perception, planning and control modules. A hybrid Transformer with shared mixed attention folds visual observations, language instructions, force and tactile signals, and policy trajectories into one latent space.

The model supports an understand-reason-execute-reflect cycle. It breaks long-horizon tasks into steps, simulates multiple physical trajectories, ranks them by predicted success and cost, executes, then evaluates the outcome. In one internal demonstration a robot switched from a three-finger to a four-finger grasp after a failed attempt and finished the task without human restart.

1. Understand: split a long-horizon task into ordered steps from language and scene context
2. Reason: simulate several physical trajectories and rank them by predicted success and cost
3. Execute: run the chosen trajectory on the robot with force and tactile feedback in the loop
4. Reflect: score the outcome and, on failure, revise the grasp or path without a human restart

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 said that at the WAIC forum while introducing the system. Kairos runs on the robot, not a remote datacenter. An 8B variant reached 125-millisecond inference latency on NVIDIA Jetson Thor at BF16 precision with the company’s KairosRT engine. Computer Weekly reported a four-billion-parameter on-device version adapted beyond Nvidia silicon so buyers avoid single-vendor lock-in. Cloud links remain for updates, fleet coordination and rare edge cases.

Keeping inference local is a direct response to the physical-world risk he described. A round trip to a remote datacenter adds delay and a network dependency that warehouse aisles and hotel corridors cannot always guarantee. Low-latency on-device timing lets the reflect step fire before a dropped item or a blocked path becomes a safety event.

Spatial understanding draws on ACE-BRAIN-0.5, which the company says posted leading results across 12 public evaluations as of July 2026. Kairos-HomeWorld generates whole-home scenes from 300,000 Chinese residential floor plans, 5,000 simulated environments and 8,700 3D assets. The geographic calibration is explicit: layouts reflect common Chinese apartments.

Three Solutions Already in Commercial Sites

ACE Robotics paired the model with three named solutions now operating with customers.

Xiaoman integrates hardware, world-model capabilities, product data, mapping and order systems for flash warehouses and convenience formats. The company says lightweight rollout can finish in days. Full commercial targets and partner lists appear in the full WAIC 2026 product launch details.

The retail path is the most aggressive on site count. Hitting 1,000 live locations in a year, then an order of magnitude more the year after, depends on that days-scale rollout claim holding once integrators leave the pilot stores. Force control within 1 N and sub-2:1 robot-to-payload ratios are the hardware bounds the company publishes for those narrow aisles.

Cloud partners include Baidu AI Cloud, Alibaba Cloud, Huawei Cloud, Tencent Cloud and SenseCore. A Caohejing Development Zone collaboration aims at an embodied AI innovation platform. Earlier, SenseTime announced Wang Xiaogang appointed ACE Robotics chairman as the firm spun deeper into physical intelligence.

Cost and the Extreme Long Tail Still Dominate

ACE Robotics itself lists the hardest unsolved problem as cost. “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,” the company said. “It is as much an industrial problem as a technical one.”

For mainstream controlled sites the sim-to-real gap has narrowed. What remains is “the extreme long tail of open, unstructured environments.” Reliability figures beyond the named pilots have not been published independently. Benchmarks cited by ACE are company-reported or self-submitted; third-party production audits are still scarce across the sector.

Omdia expects the Asia-Pacific embodied AI market to reach $1.5 billion in 2026. Chief analyst Lian Jye Su described the technology as still in the hype phase, with large-scale deployments largely limited to consumer rather than enterprise settings. Chinese residential training data and local cloud partners raise adaptation costs for buyers outside that ecosystem.

• Cost stack: hardware, compute and deployment operations must all clear a buyer-absorbable threshold
• Controlled sites: sim-to-real gap described as narrowed for mainstream indoor formats
• Open sites: extreme long tail of unstructured environments remains the unsolved reliability case
• Evidence gap: independent production audits still scarce; cited benchmarks are company-reported

What We Know

• Wearable capture removes robots from the data-collection step and claims order-of-magnitude throughput gains
• Named retail, laundry and outdoor deployments are live with Chinese customers
• On-device inference at low latency is demonstrated on Jetson-class hardware

What’s Unconfirmed

• Independent reliability metrics for Kairos 3.1 under continuous production loads
• Whether 1,000- and 10,000-site retail targets will be met on schedule
• How readily Chinese-layout world models transfer to other building codes and object inventories

HomeWorld Scenes Follow Local Floor Plans

Kairos-HomeWorld builds whole-home scenes from 300,000 Chinese residential floor plans, 5,000 simulated environments and 8,700 3D assets. The company is explicit that layouts mirror common Chinese apartments rather than a generic global housing stock.

That choice sharpens performance inside the domestic pilot base. Aisles, door swings, counter heights and storage patterns in the training distribution match the stores and hotels already running Xiaoman and Xiaoxin. The same calibration raises transfer cost for buyers whose building codes, shelf standards and object inventories sit outside that distribution.

Cloud partners reinforce the pattern. Baidu AI Cloud, Alibaba Cloud, Huawei Cloud, Tencent Cloud and SenseCore anchor updates, fleet coordination and edge-case handling inside a regional stack. Operators already tied to those platforms face a shorter integration path than buyers who must re-host models, remap homes and re-validate safety cases on different infrastructure.

Omdia’s view that large-scale deployments remain largely consumer rather than enterprise fits this geography. Until world-model assets and cloud hooks travel more easily, the commercial center of gravity stays where the floor plans and labor pools already align with the capture engine.

Why On-Device Timing Matters On Live Floors

Observers on X stressed that the closed understand-reason-execute-reflect loop is still missing from many perception-heavy world models, and that 125-millisecond on-device timing can matter more than raw parameter count once a robot shares space with people. The internal grasp-switch demo is the concrete illustration: after a three-finger attempt failed, the system revised to a four-finger grasp and finished without a human restart.

An 8B variant hit that latency target on NVIDIA Jetson Thor at BF16 with the KairosRT engine. A separate four-billion-parameter build runs beyond Nvidia silicon, giving buyers a path that avoids single-vendor lock-in while still keeping inference on the robot. Cloud links stay available for model updates, fleet coordination and rare edge cases that exceed local context.

The physical-world warning Wang delivered at WAIC sits behind that design. A flawed image wastes pixels; a flawed grasp can damage goods or injure a bystander. Local reflect cycles shrink the window between error and correction. That is the operational case for spending parameter budget on latency and sensor fusion rather than on ever-larger remote models alone.

Workers Become the New Data Factories

The second-order effect is labor. Mass capture turns ordinary warehouse pickers, laundry staff and home participants into high-density data sources. Parallel efforts, such as large citizen-wearable programs in China, point the same direction: scale comes from people performing everyday work rather than specialist teleoperators.

That model creates new questions around consent, compensation, data ownership and privacy once thousands of workers wear sensors daily. It also favors operators who already control large labor pools and physical sites. Smaller robotics startups without those advantages, including recent seed-stage players such as the team behind another Southeast Asian robotics seed raise, face a steeper path to comparable data volume.

On X, observers noted that the closed understand-reason-execute-reflect loop is the piece many perception-heavy world models still lack, and that 125-millisecond on-device timing matters more than parameter count for real floors. Others highlighted the paper’s finding that fixed third-person views sometimes reconstruct hand trajectories more accurately than moving egocentric cameras once head motion compounds error.

ACE Robotics is roughly 18 months into commercial life. Its data engine, unified model and named deployments form a coherent stack. Whether the industrial cost curve and long-tail robustness follow the collection breakthrough will decide if the tens of millions of hours ever translate into robots that stay useful outside carefully mapped Chinese stores and hotels.