Scaling Embodied AI Data: Qunce's UMI Pipeline on MaxCompute Serverless

Background: Embodied Intelligence Data at Scale

Qunce Intelligence (穹彻智能), founded in 2023, focuses on embodied intelligence pre-training models. They collect real-world human operation data using proprietary devices (RoboPocket) and organize video and motion trajectories into training samples. In August 2026 they released the Noe-0 embodied world model research preview. According to the Noe-0 Research Preview, cumulative collection reached 97,387.7 hours across 47 cities and 18 provincial regions, with over 2,000 collectors. Latest figures show 300,000 hours covering 70 cities.

Their approach is called "embodiment-free crowdsourced collection" . Collectors take portable devices into homes, stores, and other real scenes without teleoperating a robot arm. Using UMI (Universal Manipulation Interface), a collector demonstrates grasping and placing with a handheld gripper while the device records images and motion trajectories.

Data Characteristics and Volume

A single episode contains three video streams (head, left wrist, right wrist) that must be decoded and aligned by timestamp. Depth and pose data keep their original capture frame rates, which differ from video rates. Hand state and SLAM signals are recorded at their own frequencies. Metadata such as task, scene, device ID, and quality-inspection results must also be linked.

Daily raw data written to the data lake averages 10 TB with a peak of 50 TB , implying ~3.6 PB/year. Before delivery to training, data passes through decoding, timestamp alignment, re-encoding, quality inspection, and metadata extraction. When processing rules change, the same batch may need re-processing.

Data Organization: Lance Columnar Format

Data is organized by episode and stored in Lance columnar format. The camera table uses the aligned timeline with one row per frame. Other tables (depth, pose, metadata) retain original capture frames; cross-table joins use original frame numbers. This design allows:

Adding new metadata columns by writing only new column files, without rewriting existing columns.

Adding a new sensor by creating a new table; old datasets remain readable.

Training, annotation, visualization, and batch processing all reuse the same frame mapping.

Three Core Engineering Challenges

Pulsed compute demand : Iterations change alignment rules, encoding parameters, or quality thresholds. Affected data must be recomputed, sometimes expanding to the full dataset. Baseline load is low, but concentrated re-processing may require tens of thousands of cores. Buying for peak leaves hardware idle; buying for baseline prolongs re-processing.

Deep Python ecosystem dependency : Frame extraction, H.264/AV1 codec, point-cloud processing, quality inspection, and pre-annotation model inference rely on pandas, OpenCV, PyTorch, etc. Pure SQL cannot reuse this logic; pulling massive data locally is infeasible. Some quality-inspection and pre-annotation steps call large models.

Fault tolerance necessity : A single episode can be GB-scale; a worker processing one episode may take tens of minutes. In batch runs of tens of thousands of episodes, file corruption, timeouts, or format anomalies cause individual task failures. The system must record failure reasons per episode, let other tasks continue, and retry only failed episodes.

Why Not Self-Built Infrastructure?

For a startup, self-building big-data infrastructure incurs three cost categories:

Hardware & data center: racks, power, cooling, cross-data-center lines, capacity planning, idle cost, hardware refresh.

Software & operations: scheduler, storage, compute engine selection and integration; version upgrades, patches, capacity planning; monitoring, alerting, on-call, incident response; permissions, auditing, data-security compliance.

Talent: dedicated staff for platform ops and data security, plus support for algorithm engineers debugging distributed jobs.

Solution: MaxCompute Serverless + MaxFrame

Elastic Compute for Pulsed Workloads

Qunce uses MaxCompute Serverless elastic compute. At the time of the presentation, the pipeline had a 100,000 CPU core quota . When alignment rules change, affected episodes are batched and processed in parallel, drastically reducing re-processing time. Serverless lets the team request compute at job start and release it on completion, avoiding a permanently provisioned peak cluster.

Heterogeneous Compute: CPU, GPU, and Model APIs in One Pipeline

Video decoding and alignment run on CPU. For model inference (feature extraction, quality inspection, pre-annotation), the team uses GPU for custom models and calls Alibaba Cloud Bailian model services (e.g., Qwen) via API. In MaxFrame, CPU operators consume CU (compute units), GPU operators consume GU (GPU units), and model calls are metered by Token. This allows data processing and model inference to be orchestrated in a single workflow .

Job Scheduling and Custom Runtime Images

Qunce's data management platform DM3 handles governance and delivery; batch jobs run on MaxCompute. Algorithm engineers write operators using MaxFrame's pandas-style API. The platform splits and schedules jobs, distributing logic to multiple workers.

Three custom DPE (Data Processing Engine) images are maintained:

General processing image

Model inference image

Lance-format-specific image

Each image contains the dependencies required by its operators. The unit of work is one episode. Using apply_chunk, Python operators are dispatched to workers. With proper permissions, source and destination OSS buckets are mounted via ossfs2 on each worker, allowing operators to read/write OSS via local paths without embedding secrets in code. Parallelism is set by the day's valid episode count; episodes are sorted by size and round-robin assigned to workers to spread large episodes evenly.

Fault Recovery: Six-Stage Pipeline + Audit Tables

The pipeline has six stages:

Preprocessing into data lake

Format conversion

Frame extraction & re-encoding

Quality inspection

Metadata extraction

Dataset freeze

Two safeguards ensure reliability:

Daily partition overwrite : Each day's data writes to a partition dt=<date>. If a day's data fails, only that partition is re-run, avoiding duplicate appends on retry.

Per-episode audit table : Records processing result, duration, artifact path, and error message. On retry, successful episodes are reused; only failed ones are recomputed. If inputs or rules changed, affected episodes are recomputed. The audit table also provides full lineage for any episode.

MaxCompute vs. Self-Built Cluster: Responsibility Comparison

Infrastructure : MaxCompute – on-demand cloud cluster; Self-built – plan servers, network, power, cooling.

Compute Scheduling : MaxCompute – 100k CPU core quota provided; Self-built – capacity planned per load, trade-off between idle and wait time.

Tech Stack : MaxCompute – MaxFrame pandas-style API; Self-built – choose distributed framework, integrate and debug operators.

Resource Types : MaxCompute – CU/GU, on-demand Bailian model calls; Self-built – orchestrate CPU, GPU, model APIs manually.

Fault Tolerance : MaxCompute – re-run by date partition, per-episode retry via audit table; Self-built – configure retries, state tracking, recovery mechanisms.

Talent Investment : MaxCompute – team maintains processing logic, cloud handles cluster ops; Self-built – team also owns cluster ops, reliability, data security.

For Qunce's UMI data pipeline, cloud batch processing lowered the entry barrier: no need to build a peak cluster before writing and running processing jobs. When alignment rules change, concentrated compute can re-process affected data; when only a few episodes fail, targeted retries are driven by audit records.