- Client Requirement Interfacing & Specification Definition
Communicate directly with Embodied AI and World Model researchers to thoroughly analyze their large models' architectural requirements for input data (e.g., perception camera FOV, LiDAR precision, robotic arm dynamic constraints, and specific action trajectory formats).
Translate ambiguous business and scientific research pain points into high-precision "Dataset Specifications" (Specs), explicitly defining data dimensions, sensor parameters, control command action spaces, and metadata formats.
- Data Distribution Planning & Scene Design
Plan and organize the distribution of target datasets, balancing routine scenarios with long-tail scenarios (corner cases) to ensure the datasets possess strong generalizability and cover Out-Of-Distribution (OOD) extreme operating conditions.
Analyze historical or public datasets provided by clients to identify data gaps in geography, lighting, action types, and obstacle distribution, and perform targeted data completion.
- Data Production & Synthesis Orchestration
Utilize and extend Maxinsights' 3D simulation engines (such as Isaac Sim, MuJoCo, etc.) or generative world model tools to orchestrate and run large-scale data synthesis pipelines.
Write efficient scripts (Python/Bash) to automatically generate customized robot trajectories, multi-angle perception video streams, 3D point clouds, and dynamics states.
- Data Curation & Quality Governance
Responsible for data curation, filtering, and calibration, eliminating invalid data such as physical collision errors, transient sensor disconnects, or action drifts.
Write automated QA scripts to perform static and dynamic quality inspections on tens of millions of data frames across dimensions including dynamics reachability, temporal alignment, and label accuracy.
Oversee the final format packaging of datasets (e.g., converting trajectory data into MCAP, HDF5, or LeRobot formats) to guarantee seamless, direct ingestion into training pipelines.
- Closed-Loop Delivery & Continuous Cross-Team Alignment
Take ownership of data delivery and follow up to evaluate data efficacy during the early stages of client model training, establishing a closed-loop "data-to-model performance" feedback mechanism.
Maintain daily, high-frequency communication internally with the Algorithm R&D and Platform Development teams, abstracting system-level bugs or shared requirements discovered during frontline deployment to drive the standardized upgrade of Maxinsights' core data engine.
