Robotics Research Engineer

Abaka AI•Mountain View, CA

33d•$100,000 - $140,000

About The Position

Abaka AI is seeking a Robotics Research Engineer to join its Abaka Robotics Lab. This in-house research group is dedicated to creating the data foundation for physical AI. The role involves hands-on research, combining robot learning, data production, and the creation of publishable research. The engineer will train and evaluate policies on simulated and real-world robotic systems, analyze experimental results to determine data importance, and design scalable pipelines to convert raw sensor data into high-quality training datasets with minimal human input. The work will involve hardware such as 6-DoF collaborative robot arms and an ALOHA-style bimanual platform with full teleoperation. Findings from this role will directly influence the datasets developed for Abaka AI's partners and contribute to publications in leading robotics, computer vision, and machine learning conferences.

Requirements

MS or PhD in Robotics, Computer Science, Machine Learning, or a related field.
1 - 3 years of hands-on experience with real robot hardware, through research, an internship, or a project.
Strong machine learning fundamentals: optimization, generalization, evaluation methodology, and end-to-end model training.
Strong robotics fundamentals: coordinate frames, kinematics, basic control.
Strong 3D vision fundamentals: camera models, multi-view geometry, point clouds.
Rigor in evaluation: claims about a policy or a dataset are backed by measurements.

Nice To Haves

Publications or workshop papers at top-tier robotics (CoRL, RSS, ICRA, IROS), vision (CVPR, ICCV, ECCV, 3DV), or ML (NeurIPS, ICML, ICLR) venues. A widely used open-source repository or a dataset adopted by other teams is valued equally.
Experience with imitation learning or VLA codebases such as LeRobot, ACT, Diffusion Policy, OpenVLA, or π0.
Experience deploying a learned policy on a real robot and running a rigorous evaluation.
Practical 3D reconstruction experience: COLMAP, SLAM, 3DGS/NeRF, point-cloud registration.
World models or video generative models, particularly as simulators or data generators.
RL fine-tuning of learned policies, or model-based RL.
Hand and object pose estimation, hand-object interaction, or human-to-robot motion retargeting.
Depth in a simulator (MuJoCo, Isaac Sim/Lab, Genesis, SAPIEN), including environment and asset authoring.
Domain randomization, procedural scene generation, or sim-to-real transfer.
Teleoperation or data collection systems in the ALOHA / UMI / hand-tracking lineage.
Familiarity with egocentric datasets such as Ego-Exo4D, EgoMimic, HOT3D, Project Aria, and DexCap.
Dexterous or bimanual manipulation research.
Camera calibration and multi-sensor time synchronization.
Vision-language models applied to robotics: task planning, language-conditioned policies, reward or success detection.
Auto-labeling or ground-truth generation over real sensor data, with quality metrics attached.
Tactile, force-torque, or other multimodal sensing for manipulation.
Data curation and dataset-scaling studies: deduplication, mixture ratios, difficulty or diversity scoring, and ablations that identify which data mattered.

Responsibilities

Train and evaluate robot policies (imitation learning, diffusion policies, VLA, RL fine-tuning) on our data and public baselines, in simulation and on real robots.
Turn evaluation results into a quality signal for the data: which data helped, which did not, and why.
Track the state of the art: reproduce the papers that matter, run current methods on our hardware, and develop findings that hold up into publications at top robotics, vision, or ML venues.
Maintain the training and evaluation infrastructure: training runs, evaluation harnesses, and the simulation or world-model environments used ahead of hardware tests.
Design pipelines that turn raw capture (egocentric human video, robot logs, teleoperation, 3D scans) into training data with minimal human labeling: synthesize what was not captured, auto-label with models and geometry, and curate what goes into training.
Track which data decisions changed policy performance.