AI Robotics Researcher Intern - Dexterous Manipulation

Project Scope

• Learning from Human Demonstrations: Develop and refine scalable frameworks for the transfer of human-object interaction skills to diverse robotic embodiments.
• Large-Scale Data Synthesis: Architect autonomous pipelines to process vast amounts of visual data and human glove-collected data, extracting the spatial and contact-rich information necessary for generalist robot training.
• Generative Embodied AI: Implement state-of-the-art generative architectures to synthesize physically grounded, high-fidelity trajectories based on human reference motions.
• Unified Policy Training: Explore cross-embodiment representations that enable joint training on human and robot data to improve generalization in unstructured environments.
• Sim-to-Real Deployment: Research and optimize distillation and retargeting techniques to bridge the gap between simulation-trained policies and physical robotic deployment.
• Semantic Scene Understanding: Utilize vision-language foundation models to autonomously segment skills and extract task-relevant parameters from complex human activities.

Deliverables (End of Internship)

• A robust pipeline for converting human multi-modal data into actionable robot motor skills.
• A successful sim-to-real validation of a dexterous manipulation policy on a physical humanoid or multi-fingered platform.
• A high-quality technical manuscript or demo suitable for internal review or submission to a top-tier robotics conference.

Qualifications

• Master’s or Ph.D. in Robotics, Computer Science, Artificial Intelligence, Mechanical/Electrical Engineering, or related fields.
• Strong technical foundation in robot learning and control, including areas such as reinforcement learning, imitation learning, world modeling, or representation learning for agent-environment interactions.
• Practical experience implementing and fine-tuning Generative Models and Transformer architectures.
• Hands-on experience with robotic manipulation systems, particularly involving contact-rich interaction, grasping, or multi-sensor perception (e.g., tactile, force/torque, proprioception).
• Proficiency in Python and modern ML frameworks (e.g., PyTorch, JAX, TensorFlow), with experience using robotics middleware or simulation tools (e.g., ROS/ROS2, MuJoCo, Isaac Sim, PyBullet).
• Demonstrated ability to implement, experiment, and iterate on research ideas, including evaluating methods through empirical results on simulated or physical robotic systems.
• Strong analytical and system-building skills, with the ability to work across simulation, learning, perception, control, and real robot deployment as part of a larger technical team.

Preferred Qualifications

• Ph.D. (or Ph.D. candidate expecting graduation within 6–12 months).
• Prior experience with dexterous manipulation, multi-finger robotic hands, in-hand manipulation, or grasp optimization beyond parallel-jaw grasping.
• Experience deploying learning-based policies on real robotic hardware, including exposure to sim-to-real transfer challenges such as contact mismatch, compliance, sensing noise, or latency.
• Familiarity with contact modeling, tactile sensing, force/torque feedback, or low-level control interfaces for manipulation.
• Background in 3D perception, geometric representations, or learned representations relevant to physical interaction.
• Experience with reinforcement learning in continuous control, model-based methods, or real-time policy execution.
• A strong interest in building robust, real-world robotic systems, and motivation to see research ideas validated through physical experiments rather than simulation alone.
• Track record of publications in top AI or robotics conferences (CoRL, ICRA, IROS, RSS, NeurIPS, CVPR, ICML).