Bedrock

Dear Bedrock,

I’m applying for the Robotics Testing: Sim to Real Gap internship at Bedrock Robotics. I’m an EECS student at UC Berkeley with hands-on experience building data-driven pipelines, running structured evaluations, and turning noisy real-world signals into models and tools that support decisions. I’m specifically interested in simulation fidelity work where the output is measurable: defined test plans, quantified gaps, and validated improvements to a physics model.

In prior work, I built systems that start from imperfect data and end with a usable, verifiable product. For example, I created an OCR+NLP semantic search pipeline over 1600s–1800s shipping manifests, including embedding-based retrieval and automated entity extraction across noisy scans—work that required careful error analysis, calibration, and iterative validation against real documents. I’ve also worked on satellite-imagery ML at scale (10k+ tiles) to classify and detect targets for field deployment planning, where success depended on controlled data collection, robust evaluation metrics, and closing the loop between predictions and real constraints.

I also have hardware-adjacent testing experience through high-power rocketry avionics work, including building and testing custom PCB-based systems under tight mass/volume constraints and working with microcontrollers (ESP32) and instrumentation. That background maps directly onto sim-to-real characterization: designing repeatable experiments, collecting clean ground-truth data, and documenting results so simulation and ML engineers can translate findings into model changes.

At Bedrock, I would expect to contribute by:

  • Designing structured on-robot test matrices across operating conditions/configurations and defining pass/fail and fidelity metrics.
  • Building Python tooling to ingest logs, synchronize sensors, compute derived dynamics features, and generate repeatable reports.
  • Identifying the highest-leverage mismatch modes (e.g., contact/terrain interaction, actuator dynamics, latency/noise), proposing targeted data collection, and validating that model changes reduce the error on held-out scenarios.
  • Writing clear documentation that makes experiments reproducible and decisions auditable.

I can work full-time in your San Francisco office during Summer 2026. Thank you for your consideration—I’d welcome the chance to discuss how my background in rigorous evaluation and applied modeling can support Bedrock’s simulation fidelity and validation efforts.

Sincerely, Roshan Taneja