Jack Benarroch Jedlicki

I'm a second-year Computer Science PhD student at Harvard University, working on embodied intelligence, supported by the “la Caixa” Foundation Fellowship and the Harvard Prize Fellowship.

Previously, I earned two bachelor's degrees in Mathematics and Physics at the University of Barcelona. During that time, I worked on computer vision and digital twins at the Alaa Lab at UC Berkeley, on topology-informed generative models at the Computer Vision Center, and on tensor networks for efficient physics simulations at the Barcelona Supercomputing Center. Before that, I also worked on sensing and imaging techniques at MIT's Auto-ID Lab.

Email  /  CV  /  Scholar  /  GitHub  /  LinkedIn

Research

I’m interested in embodied AI, especially robot manipulation, memory, and long-horizon reasoning.

We turn GPT-6-Astra into an embodied agent by connecting it to a simulated Franka arm through a simple harness. On RoboMME’s 16 memory-dependent tasks, it reaches a 92.1% success rate, surpassing human performance (90.5%) and well above the best leaderboard score (60.8%).

We show that a hybrid design, keeping short-term perceptual memory inside the policy and moving long-term object memory to an external, readable object ledger, reaches a 64.3% average on RoboMME, well above the previous state of the art (60.8%).

We design a memory-aware embodied agent following Occam’s razor: remove memory from the policy and give it only to the high-level planner. Our system outperforms memory-conditioned policies and VLMs that reason over raw frame history in both simulation and real-world experiments.

Distributional reward models that express uncertainty, showing that common ensemble aggregation rules (mean, worst-case, uncertainty-weighted) are special cases of a single expression.

Generation of digital twins of the cardiovascular system using computer vision and physics-informed neural networks.

Introduces new methods for training generative models via differentiable topological regularizers, with novel proofs of differentiability and new connections between topology and generative modeling.

Study and analysis of tensor-network–based quantum-inspired simulators, demonstrating empirically that they can outperform classical algorithms in accuracy while using significantly less memory, enabling the solution of memory-intensive problems without quantum hardware.

Sensing and imaging techniques for metal particle detection in freeze-dried vaccines.

Additional Projects

Some projects done in courses @ Harvard or MIT.

Experimental study of the generalization and scaling behavior of diffusion policies under distribution shifts, including novel object configurations and obstacle-rich environments.

Theoretical and experimental comparison of generative modeling approaches—including diffusion models, flow matching, and rectified flows—across image generation and robotic action generation tasks. We show that flow matching, unlike diffusion, enables one-shot inference for robot action trajectory generation.