Amal Mehta

Independent ML & Robotics Research

San Francisco Bay Area

Building and evaluating learning systems end to end: decoding hand movement from motor-cortex spikes, post-training a small transformer with GRPO, testing how well RL and meta-learned drone policies transfer from simulation, and training a Go engine from scratch by self-play. Some of these projects are featured above, and the rest are below. Also built tools for researchers, such as Paper to Playground. Member of the Sensorimotor AI journal club at Berkeley’s Redwood Center for Theoretical Neuroscience, and mentor to undergraduate scholars from UC Berkeley.

Neural decoding GRPO Sim-to-real Meta-learning Self-play RL

Deep Learning Scientist · PhysicsAI

DARPA-funded · remote

Turned synthetic-data generation into a closed RL loop: agents (TD3, PPO, DDPG) set an image generator’s parameters, a spatiotemporal object detector retrained on what it produced, and the detector’s accuracy on real drone imagery came back as the reward. Real-world performance, not hand-tuned rendering, drove every iteration and closed the sim-to-real gap.

Owned the meta-RL codebase: a PEARL agent over TD3 that learns a probabilistic embedding of each generator task, so it adapts to a new one in a few tries, and that was also applied to a fighter-jet simulation. Ran and monitored its multi-week GPU training runs.

Also implemented domain adaptation from scratch to hold the detector’s accuracy with scarce real labels and mostly synthetic data (10–15 GB datasets), and built and benchmarked the GAN pipelines (DCGAN, InfoGAN, FineGAN) that produced its training images.

TD3 PPO DDPG PEARL meta-RL Domain adaptation GANs

Award: team recognized as the top group in the nation for its reinforcement-learning work

Deep RL Researcher · Hybrid Robotics Lab

UC Berkeley · Prof. Koushil Sreenath

Trained one RL policy that takes a Unitree Go1 robot dog through the whole task (walk to the ball, kick it at a target, stay balanced) in place of a separate policy per skill. On the real robot it hit the target about 60% of the time: it learned entirely in simulation, which can’t fully reproduce the ball’s friction on the ground, how it deforms when kicked, or sensor noise. Took graduate classes in RL and control theory alongside.

The policy is PPO, trained in IsaacGym on a three-stage reward: reach the ball, kick it, then wait to see whether it scores. A reward gradient, a stage flag and masked inputs (the ball and goal positions are zeroed until they matter) let one network move between stages smoothly, and CPG-based gait tracking kept its walking steady. Also prototyped RLPD, SAC that learns from prior data, to fine-tune the simulation-trained policy on real robot runs and narrow the gap to hardware.

PPO IsaacGym Multi-stage rewards Sim-to-real RLPD ROS OpenCV RealSense

Report (PDF, 330 KB) →

Research Assistant · Video & Image Processing Lab

BAIR, UC Berkeley · Prof. Avideh Zakhor

Built a deep-learning pipeline that outlines invasive melanoma pixel by pixel in very large tissue images, to help pathologists measure how far a cancer has spread.

Invasive and in-situ melanoma look alike under the microscope; what matters is whether the tumour has crossed below the epidermis. So the pipeline segments the epidermis and the melanoma separately, at four magnifications, and keeps the melanoma outside the epidermis. Trained on slides from 57 UCSF patients outlined by two pathologists, it raised F1 on invasive melanoma to 0.44, from 0.14 for the previous best.

Most of the work was engineering around the slides: cutting very large images into patches that fit on two GPUs, dropping empty ones, removing glue stains, and rebalancing rare classes with flips and rotations. HRNet+OCR started from Cityscapes weights and was benchmarked against HookNet, U-Net and the earlier published model.

Published: SPIE Medical Imaging 2023 · USCAP 2022, Stowell-Orbison Award Poster Session

Paper (SPIE Digital Library) →

Researcher · Abbasi-Asl Lab, UCSF

UC San Francisco · Prof. Reza Abbasi-Asl

Built pipelines that line up 2D and 3D brain scans so they can be compared, such as the same brain across visits or different brains against a common template. They rest on empirical mode decomposition, which splits an image into layers from fine detail down to a smooth residue, extended here from flat signals to full 3D volumes. Methods from the lab later went into the Neuroimaging Platform Mac app.

Brain image registration Signal decomposition

Deep Learning Segmentation for Invasive Melanoma

SPIE Medical Imaging · peer-reviewed paper · with A. Shah, M. Wang, N. Neumann, T. McCalmont and A. Zakhor

Paper (SPIE Digital Library) →

Quantifying Invasive Melanoma Volume by Deep Learning Segmentation at the Pixel Level

USCAP Annual Meeting · abstract, Stowell-Orbison Award Poster Session · with N. Neumann, M. Wang and others

Robot soccer with deep RL

Research project report · Hybrid Robotics Lab, UC Berkeley · advisor Prof. Koushil Sreenath

Report (PDF, 330 KB) →

Few-Shot Imitation Learning for Robotic Manipulation

Class report · CS330 Meta-Learning, Stanford · with Sebastian Kochman and Soham Sinha

Teaching a robot arm a new Meta-World task from a single expert demonstration. We paired inverse soft Q-learning with implicit MAML, to our knowledge the first Q-learning inside iMAML, and tried a recurrent RL² variant. Both were compared with a MAML baseline trained by behavioural cloning.

Meta-learning Imitation learning iMAML RL² Meta-World

Report (PDF, 445 KB) →

Defensive Iterative Retraining

Class report · CS285 Deep Reinforcement Learning, UC Berkeley · with Aman Shah and Srikar Boinapally

Training agents to resist adversarial policies by mixing past selves and freshly trained adversaries into the opponent pool, instead of plain self-play. In SumoAnts it slightly beat self-play; in the harder SumoHumans it fell behind, because the interruptions came before the agent had learned to stay on its feet.

Multi-agent RL Adversarial policies PPO

Report (PDF, 200 KB) →