Jade Choghari

I'm currently working on Physical AI research at Scale AI, focusing on robot foundation models, vision-language understanding, and long-horizon robot learning.

My recent work includes FineART-VLA, one of the first open-source System 1 / System 2 vision-language-action models, where a high-level model predicts language subtasks to guide low-level robot actions. I also work on large-scale video understanding and real-world evaluations of robot foundation models.

Previously, I was part of the founding team behind LeRobot at Hugging Face 🤗, an open-source library for end-to-end robot learning with nearly 29K GitHub stars, published at ICLR 2026. Before robotics, I contributed to computer vision and multimodal AI.

I studied Computer Science at the University of Waterloo.

Email  /  CV  /  Scholar  /  Twitter  /  LinkedIn  /  Github

profile photo

Research

I'm interested in robot learning, large-scale simulation, and vision-language-action (VLA) policies. My research focuses on building scalable environments, datasets, and transformer-based control systems that enable robots to perform complex manipulation and real-world tasks. Below is a selection of my research and open-source work.

Research/Projects Highlights

Augmentron paper figure of multi-view robot manipulation environments Augmentron: Scalable Multi-View Visual Augmentation for Robot Learning
Tanay Tandon, Aseem Doriwala, Jade Choghari, Catherine Weaver, Pragna Mannam
8th Robot Learning Workshop, NeurIPS 2026
Paper

Introduces Augmentron, a geometry-aware method for generating temporally consistent, multi-view visual augmentations of robot demonstrations. Improves policy generalization to new distractors and workspace layouts across three manipulation tasks.

FineART paper figure showing robot tasks and dataset composition FineART: Fine-Grained Annotated Robotic Trajectory Dataset and Vision-Language-Action Model for Bimanual Manipulation
Jade Choghari, Pepijn Kooijmans, Mansi Agarwal, Yusuf Umut Ciftci, Aseem Doriwala, Catherine Weaver, Mouli Sivapurapu, Kai Yang, Thomas Wolf, Jackson Lee, Pragna Mannam
arXiv, 2026
Paper / Blog / Dataset / Model

Introduces FineART, a dataset of 40,543 bimanual episodes with 533,913 labeled subtasks, and FineART-VLA, one of the world's first open-source System 1 / System 2 VLAs. System 2 predicts the next subtask in language, which guides System 1's continuous robot actions.

Scale AI Physical AI photo of a robot-data collection setup Scale AI's Physical AI Evaluation-as-a-Service
Launched October 2026
Website

Real-world evaluation of robot policies across embodiments and tasks, including spatial understanding, long-horizon manipulation, and failure recovery. Part of the Scale AI team that built this offering from the ground up and launched it in October 2026.

Scale AI Physical AI image illustrating multimodal robot data and annotation The Path to Large Scale Dense Video Captioning
Jade Choghari, Agustin Sansone, Nicolas Pasqualis, Conrado Mader, Aleks Tiupikov, Mouli Sivapurapu
Scale Labs, May 2026
Blog

An empirical study of vision-language models for captioning fine-grained manipulation actions. Hand-cropped past/present/future frames improved caption acceptance from 28.7% with raw video to 61.0%.

LeRobot ICLR paper figure showing its robotics data and model training stack LeRobot: An Open-Source Library for End-to-End Robot Learning
Rémi Cadène, Simon Aliberts, Francesco Capuano, Michel Aractingi, Adil Zouitine, Pepijn Kooijmans, Jade Choghari, Martino Russi, Caroline Pascal, Steven Palma, Mustafa Shukor, Jess Moss, Alexander Soare, Dana Aubakirova, Quentin Lhoest, Quentin Gallouédec, Thomas Wolf
ICLR, 2026
Paper / Code

Introduces LeRobot, an open-source library for real-world robot learning that integrates hardware control, datasets, policy training, and model evaluation.

X-VLA: Soft-Prompted Vision-Language-Action Foundation Model
2025

Built X-VLA in LeRobot, the first soft-prompted VLA model capable of scaling across many embodiments, cameras, action spaces, and environments through a unified transformer backbone. We release 6 checkpoints, including a cloth-folding model achieving 100% success over a 2-hour continuous run, with a 1.5k-episode cloth-folding dataset to support community fine-tuning. Docs/Blog, PR, Twitter

EnvHub: A Community Push to Scale Simulation Environments
2025

Launched EnvHub, a large-scale initiative to make simulation environments shareable and reusable across the community. Enables one-line loading of Isaac, MuJoCo, Genesis, and custom tasks into LeRobot—reviving the 2017 OpenAI Gym call-to-action, but for modern robot learning. Docs, Twitter

LeRobot: Unified Evaluation Stack for VLA Benchmarks
2025

Built the evaluation stack for LeRobot, integrating LIBERO, MetaWorld, and other benchmarks to evaluate VLA models across 130+ manipulation tasks. Github, Twitter

VLAb: Pretraining Toolkit for VLA Models
2025

A streamlined library for pretraining VLA models with multi-dataset support and distributed training. Built the pretraining stack used to reproduce SmolVLA and scale workflows across multi-GPU and SLURM clusters. Github

LeRobot: Machine Learning Framework for Real-World Robotics
2025

An open-source PyTorch library for end-to-end robot learning, providing datasets, policies, simulation tools, and training pipelines across diverse robots. Github

Gym-Genesis: GPU-Accelerated Simulation Environments for Robotics
2025

A vectorized Gym-style environment wrapper for the Genesis physics engine, enabling thousands of parallel environments on GPU for high-throughput robot learning. I developed core environment wrappers, improved observation pipelines, and added SO101 with imitation-learning demos. Github

Roomi Robot: Open-Source Autonomous Housekeeping Robot
2025

A low-cost mobile manipulation robot for housekeeping tasks, combining a mobile base, dual arms, and multi-camera perception for tasks like towel replacement, trash collection, and restocking. Github

SmolVLA: A Vision-Language-Action Model for Efficient Robotics
2025

A lightweight VLA model designed for affordable robot learning with scalable simulation and training pipelines. I built the simulation stack, developed scalable environments, and improved throughput, realism, and reproducibility for large-scale VLA training. Github

RT-DETRv2: Improved Real-Time Detection Transformer
2025

An enhanced real-time DETR with selective multi-scale sampling, optional discrete attention for easier deployment, and dynamic data augmentation for stronger performance. I contributed the official RT-DETRv2 integration in Hugging Face Transformers, including model architecture, preprocessing, and training utilities. Hugging Face Transformers, LinkedIn

TextNet: Fast Backbone for Arbitrary-Shaped Text Detection
2023

A NAS-designed backbone for detecting arbitrarily-shaped and rotated text, using asymmetric kernels to capture extreme aspect ratios. I contributed the TextNet integration in Hugging Face Transformers and added the TextNetForImageClassification variant. Hugging Face Transformers, Twitter

Ferret-UI: Grounded Mobile UI Understanding with Multimodal LLMs
arXiv 2024

A multimodal LLM specialized for mobile UI screens, with strong referring, grounding, and reasoning abilities across fine-grained UI tasks. I integrated Ferret-UI into Hugging Face and contributed to the model and demo tooling. Hugging Face Transformers Twitter

VidToMe: Video Token Merging for Zero-Shot Video Editing
arXiv 2024

Improves zero-shot video editing by merging redundant tokens across frames to enhance temporal consistency and reduce memory usage. I integrated VidToMe into Hugging Face and helped build the public demo. Github, Twitter

Quality-Aware Masked Diffusion Transformer for Music Generation
arXiv 2024

Introduces a quality-aware training framework and masked diffusion transformer for high-fidelity text-to-music generation, achieving SOTA results on MusicCaps and Song-Describer. I integrated QA-MDT into Hugging Face and contributed to model and demo tooling. Github, Twitter

VoiceRestore: Flow-Matching Transformers for Speech Recording Quality Restoration
2024

Restores degraded speech using a unified flow-matching transformer trained on synthetic data, handling noise, reverberation, compression, and bandwidth artifacts. I integrated VoiceRestore into Hugging Face and contributed to the demo + tooling. Github, Twitter

VFusion3D: Scalable 3D Generation from Video Diffusion Models
ECCV 2024

Learns 3D assets from a single image using video diffusion models and synthetic multi-view data. I integrated VFusion3D into Hugging Face, built the full model architecture, and created the public Gradio demo. Github

GTI: A Scalable Graph-Based Trajectory Imputation Method
ACM SIGSPATIAL 2023

GTI is a scalable trajectory imputation approach that reconstructs sparse GPS trajectories without relying on existing maps, enabling data completion for map construction and urban mobility applications. Paper