Kehlani Fay 🤖

Dexterous manipulation is limited by both control and design, without consensus as to what makes manipulators best for performing dexterous tasks. This raises a fundamental challenge: how should we design and control robot manipulators that are optimized for dexterity? We present a co-design framework that learns task-specific hand morphology and complementary dexterous control policies. The framework supports 1) an expansive morphology search space including joint, finger, and palm generation, 2) scalable evaluation across the wide design space via morphology-conditioned cross-embodied control, and 3) real-world fabrication with accessible components. We evaluate the approach across multiple dexterous tasks, including in-hand rotation with simulation and real deployment. Our framework enables an end-to-end pipeline that can design, train, fabricate, and deploy a new robotic hand in under 24 hours.

@article{fay2025crossembodied,
  title={Cross Embodied Co-Design for Dexterous Hands},
  author={Fay, Kehlani and Djapri, Darin and Zorin, Anya and Clinton, James and El Lahib, Ali and Su, Hao and Tolley, Michael T. and Yi, Sha and Wang, Xiaolong},
  journal={International Conference on Learning Representations (ICLR)},
  year={2026}
}

Catching arbitrary objects requires robots to adapt in real-time to changing visual observations, unseen object properties, and diverse contact dynamics. This paper presents Catch It All!, an autonomous catching framework for humanoid robots that solves this problem without motion capture or teleoperation. Catch It All! develops a generalizable, egocentric control pipeline at 50 Hz alongside a safety-aware sim-to-real curriculum allowing real-time catching of novel objects. The policy is trained entirely in simulation using safety-aware curricula that ensure safe behavior near torque limits and avoid self-collision. The resulting policy is deployed zero-shot on a Unitree G1 humanoid robot without any fine-tuning. We evaluate across 11 novel objects spanning wide variation in size, shape, toss ranges, mass, and restitution, achieving strong catch success rates in both simulation and real-world experiments.

@misc{fay2026catchitall,
  title={Catch It All!: Generalizable Dynamic Catching with a Dexterous Hand},
  author={Fay, Kehlani and Shukla, Arth and Liu, Patrick and Djapri, Darin and Buynitsky, Alexiy and Hansen, Nicklas},
  year={2026},
  note={Pre-print, under review}
}

Cross-embodiment generalization underpins the vision of building generalist embodied agents for any robot, yet its enabling factors remain poorly understood. We investigate embodiment scaling laws, the hypothesis that increasing the number of training embodiments improves generalization to unseen ones, using robot locomotion as a test bed. We procedurally generate ~1,000 embodiments with topological, geometric, and joint-level kinematic variations, and train policies on random subsets. We observe positive scaling trends supporting the hypothesis, and find that embodiment scaling enables substantially broader generalization than data scaling on fixed embodiments. Our best policy, trained on the full dataset, transfers zero-shot to novel embodiments in simulation and the real world, including the Unitree Go2 and H1. These results represent a step toward general embodied intelligence, with relevance to adaptive control for configurable robots, morphology co-design, and beyond.

@inproceedings{ai2025embodiment,
  title={Towards Embodiment Scaling Laws in Robot Locomotion},
  author={Ai, Bo and Dai, Liu and Bohlinger, Nico and Li, Dichen and Mu, Tongzhou and Wu, Zhanxin and Fay, Kehlani and Christensen, Henrik I. and Peters, Jan and Su, Hao},
  booktitle={Conference on Robot Learning (CoRL)},
  year={2025}
}

Researchers tag and track marine animals to study migration patterns, human impacts on behavior, and behavioral shifts due to climate change. Accurate data collection often requires tagging individual animals to collect spatio-temporal state estimates of the animal's geo-position and depth. Acoustic transmitters are prominent due to their continuous communication without requiring retrieval or surfacing to collect data. These transmitters emit underwater acoustic pulses which can be detected by hydrophones. However, the frequent movement of aquatic animals results in high data loss when the animal moves out of the detection range of a stationary hydrophone. Autonomous underwater vehicle (AUV) systems offer a solution for localizing transmitters with higher resolution over longer periods of time. Such systems previously deployed have often required multiple hydrophones mounted on a large frame carried by the AUV, increasing drag and limiting speed. This work provides an alternative by equipping multiple AUVs with a single compact hydrophone payload. A particle filter algorithm equipped with a hidden Markov model (HMM) behavioral motion model fuses measurements from multiple AUVs to estimate the transmitter's position. Real-world data shows a root mean square error (RMSE) of approximately 10 meters for short-term deployments, and a larger simulated dataset shows an RMSE of approximately 15 meters for longer deployments over a larger area. The HMM fit to historical animal movement data outperforms a generic velocity motion model, and both outperform a baseline random walk motion model.

@article{herrera2026multiauv,
  title={Multi-AUV Marine Life Tracking with Single Hydrophone Payloads via a Hidden Markov Model Equipped Particle Filter},
  author={Herrera, Christopher and Fay, Kehlani and Clark, Christopher and Soto, Alberto and Lowe, Christopher G. and Espinoza, Mario},
  journal={arXiv preprint arXiv:2606.22335},
  year={2026}
}

Traditional techniques for marine life tracking use stationary receivers that detect and obtain measurements from tagged animals. Recently, such static systems have been replaced by multiple mobile robots, e.g., autonomous underwater vehicles (AUVs), equipped with omni-directional hydrophones that can accurately localize marine life. In this paper, the application of homogeneous multi-AUV systems to track and localize marine life is used as a motivating example to develop new MRMP (Multi-Robot Motion Planning) algorithms. These algorithms generate trajectories that maximize a new fitness function that incorporates 1) probabilistic motion models generated from historical data of live sharks, and 2) ideal AUV formations for observing a shark from multiple sensor vantage points. The two expansive RRT variants, named Independent State Expansion (ISE) planning and Joint State Expansion (JSE) planning, differ in how new samples are randomly generated during the algorithm's random search. The fitness function was developed to quantify how accurately the positioning of AUVs would trilaterate the target animal. Through simulation, it was found that the Joint planner was 70% faster with respect to run time than the Independent planner, while both could produce similar mean fitness function values. The fitness for these variants was also measured for simulations where different target motion models were used when calculating the fitness function, highlighting the improved performance when using actual target motion models.

@inproceedings{fay2023multirobot,
  title={Multi-Robot Expansive Planning and Trajectory Evaluation for Tracking and Localization of Marine Life},
  author={Fay, Kehlani A. and Giordano, Michael and Soto, Alberto and Clark, Christopher M. and Spurgeon, Emily and Anderson, James and Lowe, Christopher G.},
  booktitle={Proceedings of the 38th ACM/SIGAPP Symposium on Applied Computing},
  series={SAC '23},
  pages={808--815},
  year={2023},
  publisher={ACM},
  doi={10.1145/3555776.3577749}
}