Zohar Rimon
I'm an EECS PhD candidate at Technion under the supervision of Aviv Tamar. Interested in how agents "sense to act" and "act to sense" in the world.
Contact me: zohar.rimon at campus dot technion dot ac dot il
News
- — Started a VSR with Pieter Abbeel at UC Berkeley.
- — "More with LESS" accepted to RSS 2026.
- — "Task Tokens" accepted to ICLR 2026.
- — "Toward Artificial Palpation" accepted to NeurIPS 2025.
- — "MAMBA" accepted to ICLR 2024.
- — Gave a talk at Siddharth Srivastava's group, ASU, on Generalization in Meta-RL.
- — "NeRN" accepted to ICLR 2023 as a Spotlight (top 25%).
- — "Meta-RL with Finite Training Tasks" accepted to NeurIPS 2022.
More with LESS -- Local Scene Representations for Tactile Imaging
RSS 2026
The first ever hand-held, 3D, real-time, general tactile imaging system. We leverage the fact that touch is a local sensation, using a grid of local, compositional encoders instead of a global one. We generalize from single-inclusion training to multi-inclusion objects it's never seen — and it even works hand-held, no robot arm required.
Toward Artificial Palpation: Representation Learning of Touch on Soft Bodies
NeurIPS 2025
Humans learn to understand touch by playing with countless toys as kids. We do the same for robots: a tactile sensor palpates soft phantoms for hundreds of hours, learning a representation of touch from scratch. The result — a robot that detects lumps in soft tissue far more accurately than the human hand.
Task Tokens: A Flexible Approach to Adapting Behavior Foundation Models
ICLR 2026
How do you specialize a generalist robot policy without breaking what makes it generalist? Task Tokens teach a frozen behavior foundation model new tricks via a lightweight, RL-trained encoder — no fine-tuning required. The model keeps its diverse control skills while getting sharper at the task you actually care about.
MAMBA: an effective world model approach for meta-reinfocement learning
ICLR 2024
Meta-RL agents are notoriously sample-hungry and struggle past low-dimensional tasks. MAMBA fuses world models with meta-RL to learn new tasks up to 15x faster, with barely any hyperparameter tuning. It also holds up on much harder, higher-dimensional domains — a real step toward generalizing agents.
NeRN - Learning Neural Representations for Neural Networks
ICLR 2023 - Notable Top 25% (Spotlight)
What if a neural network's weights were themselves the output of another neural network? NeRN assigns each convolutional kernel a coordinate and learns a predictor that maps coordinates straight to weights — reconstructing full CNNs on CIFAR-10, CIFAR-100, and ImageNet with smoothness and distillation tricks to keep it stable.
Meta-RL with Finite Training Tasks - a Density Estimation Approach
NeurIPS 2022
How many training tasks does meta-RL actually need? We show the answer hinges on the intrinsic dimension of the task distribution, not the number of states and actions — and prove it by learning that distribution directly with kernel density estimation, yielding tighter bounds and better real-world generalization.
You Got Me Dancing
Undergrad Project - 2020
Can you drop a dancer into someone else's living-room video and make it look real? We call it Scene Aware Motion Transfer: a fully automatic pipeline that tracks individual identities in crowded, in-the-wild footage and transfers motion onto them, seamlessly integrated into the real scene.