• Own the robot-learning pipeline end to end. Establish and maintain a trusted workflow spanning robot data
collection, data ingestion and compilation, post-training, checkpoint generation, inference, and real-robot evaluation.
• Define the supported golden path. Maintain known-good combinations of code, datasets, configurations, checkpoints, robot software, hardware settings, task stations, and evaluation procedures.
• Build automated validation at every interface. Develop checks for timestamp synchronization, sensor and action integrity, episode completeness, schema compatibility, dataset migrations, dataloader outputs, preprocessing behavior, model inputs, and configuration correctness.
• Create end-to-end regression tests. Build representative smoke tests that exercise data compilation, training, checkpoint loading, inference, replay or simulation, and real-robot execution. Develop small-scale overfit and canary experiments that catch correctness regressions before expensive training runs begin.
• Ensure training and inference consistency. Identify and prevent discrepancies in image processing, sensor normalization, temporal context, action representation, model configuration, and other transformations used across training and deployment.
• Make failures observable and diagnosable. Build instrumentation and debugging tools that help determine whether a performance regression originated in data, model code, infrastructure, inference, robot software, hardware configuration, the physical environment, or evaluation execution.
• Improve real-robot evaluation reliability. Partner with researchers and robot operations to establish stable benchmark stations, reference baselines, clear rubrics, repeatable trial protocols, operator procedures, and tracking of environmental variables that affect performance.
• Lead cross-functional root-cause investigations. Drive ambiguous failures to resolution across Research, Data Infrastructure, Model Infrastructure, Software, and Robot Operations. Turn incidents and regressions into durable tests, monitors, documentation, and interface contracts.
• Establish release and compatibility standards. Define the validation required before changes to robot software, data systems, training code, or inference systems become part of the supported research workflow.
• Measure and improve research reliability. Track pipeline success rates, reproducibility, regression frequency, time to root cause, benchmark stability, and other metrics that reflect the health of the research workflow.