Eric Jang

wiki
, , , ,
1

American research engineer and roboticist known for co-inventing the Gumbel-Softmax distribution and for leading the AI development of humanoid robots at 1X Technologies.

Profile
Nationality American
Education ScB, Applied Mathematics & Computer Science; ScM, Computer Science — Brown University (2012–2016)
Previous Institution Google / Robotics at Google (Google DeepMind Robotics)
Research Areas Robotics, Imitation Learning, Reinforcement Learning, Generative Models, Foundation Models for Robotics
Personal Website evjang.com
X / Twitter @ericjang11
GitHub ericjang
Google Scholar Eric Jang

Overview

Eric Jang is an American researcher and engineer whose work spans the intersection of deep learning theory and physical robotics. He is best known as co-inventor of the Gumbel-Softmax (Concrete) distribution — a foundational technique enabling backpropagation through discrete categorical samples — published at ICLR 2017 and now ubiquitous across generative modelling and structured prediction. At Robotics at Google, he spent six years applying the scaling philosophy of large-model deep learning to physical manipulation, contributing to landmark projects including SayCan, QT-Opt, Grasp2Vec, and BC-Z. From 2022 to early 2026, he served as Vice President of AI at 1X Technologies (formerly Halodi Robotics), where he built the AI team from scratch and oversaw the development of the Redwood AI model and the 1X World Model for the NEO humanoid. He left 1X in January 2026 and is currently on a self-described sabbatical, exploring new directions in AI and robotics research.

Early Life & Education

Jang attended Brown University from 2012 to 2016, completing a concurrent ScB in Applied Mathematics and Computer Science and a ScM in Computer Science. Between 2013 and 2015, he undertook internships at Pixar, Google, and Two Sigma, experiences he later described as formative in shaping his instincts for applied research at the boundary of theory and engineering. His early blog writing on neural network topics — including detailed implementations of GANs and variational autoencoders in TensorFlow — established his public profile in the ML community while he was still an undergraduate.

Career

Robotics at Google / Google DeepMind Robotics (2016–2022)

Jang joined Google in 2016 as a Senior Research Scientist on the Robotics team, a role he held for six years. His research focused on applying the principle that rich data combined with simple learning algorithms and powerful compute outperforms hand-engineered systems — this time in the domain of robotics.

His most cited individual contribution from this period was the co-invention of the Gumbel-Softmax distribution (with Shixiang Gu and Ben Poole), a method for differentiable sampling from categorical distributions, published as Categorical Reparameterization with Gumbel-Softmax at ICLR 2017. The technique was discovered concurrently and independently by Chris Maddison, Andriy Mnih, and Yee Whye Teh at DeepMind, who named it the Concrete Distribution; both papers are now standard citations in work involving discrete latent variables.

On the robotics side, he co-led the Google Brain Moonshot team — a group of more than 20 researchers — that produced SayCan (Do As I Can, Not As I Say, 2022), a system grounding large language model reasoning in robotic affordances to enable long-horizon mobile manipulation in real kitchen environments. He also contributed to QT-Opt, a large-scale deep RL system for vision-based grasping; Grasp2Vec, which enabled instance grasping through self-supervised representation learning; Time Contrastive Networks (TCN) for cross-embodiment imitation; and BC-Z, a zero-shot task generalisation system via behaviour cloning on language-conditioned demonstrations. He also developed Tensor2Robot, a ML framework used by the robotics manipulation team and Everyday Robots.

He served as area chair for ICML, CoRL, and NeurIPS, and reviewed for multiple other venues throughout this period.

1X Technologies / Halodi Robotics (2022–2026)

A key inflection point for the company came in 2022 when Halodi began merging its humanoid hardware with modern AI approaches, hiring former Google Brain researcher Eric Jang as the first California-based employee. In March 2023, Halodi rebranded as 1X Technologies and closed a $23.5 million Series A round led by the OpenAI Startup Fund.

Jang and his first hire worked from his garage for several months to save money before the company expanded its Bay Area presence. As VP of AI, he built out the team responsible for reinforcement learning whole-body controllers, vision-language-action models (VLAs), vision-language models (VLMs), world models, simulation infrastructure, and the data engine for the EVE and NEO robots. He led the development of Redwood AI, an end-to-end onboard AI model for the NEO humanoid, which 1X described as among the first unified whole-body mobile manipulation systems for bipedal robots. He also oversaw the 1X World Model (1XWM), a generative cognitive core enabling the robot to simulate tasks before executing them.

Jang’s tenure was defined by a data-first philosophy that challenged traditional robotics paradigms, and his departure in January 2026 came at a high-water mark as 1X unveiled a major evolution of its world model.

Sabbatical (2026–present)

Following his departure from 1X, Jang entered a self-described sabbatical. He has been re-implementing deep learning papers, travelling to China to meet companies in the Chinese robotics ecosystem, and working on a large tutorial project for his blog. His current public-facing project, AutoGo, involves rebuilding AlphaGo from scratch using modern AI tools, which he presented in a long-form technical interview and accompanying tutorial. His blog post “Leaving 1X” articulated a thesis about the next generation of “magical objects” in AI — generative video models, reasoning systems — that he believes will unlock new possibilities for robotics.

Key Contributions

  • Gumbel-Softmax / Concrete Distribution — Co-invented (with Shixiang Gu and Ben Poole) a gradient estimator enabling backpropagation through categorical discrete samples, published at ICLR 2017. Concurrent with and independently confirmed by the “Concrete Distribution” paper from DeepMind. Now a standard tool for training models with discrete latent variables.

  • SayCan — Do As I Can, Not As I Say (2022) — Co-led the Brain Moonshot team of 20+ researchers that grounded large language model planning in the value functions of a real robot, enabling hundreds of long-horizon mobile manipulation tasks in kitchen environments. One of the most cited papers in LLM-robotics grounding.

  • BC-Z: Zero-Shot Task Generalisation with Robotic Imitation Learning (2021) — Demonstrated that behaviour cloning on language-conditioned teleoperated demonstrations enables generalisation to unseen task instructions; introduced key design patterns including action chunking and auxiliary prediction heads that became standard ingredients in subsequent vision-language-action models.

  • QT-Opt: Scalable Deep RL for Vision-Based Robotic Manipulation (2018) — Large-scale reinforcement learning system for grasping from monocular RGB images on real robot hardware, demonstrating emergent behaviours including object singulation and retry from failure.

  • Time Contrastive Networks (TCN) (2017) — Proposed unsupervised multi-viewpoint representation learning for robotic imitation, enabling a robot to mimic human pouring behaviour from a single demonstration video without explicit correspondence.

  • Grasp2Vec (2018) — Self-supervised instance grasping without object labelling, learning object representations through the robot’s own grasping interactions.

  • Tensor2Robot — Open-source ML framework for the Google Robotics manipulation team, standardising model training pipelines across the group.

  • Normalizing Flows Tutorials (2018) — Two-part blog tutorial introducing normalizing flows to a wide ML audience, implemented in JAX. Became one of the most widely cited informal introductions to the topic and helped popularise both the technique and JAX at the time of the latter’s early development.

  • AutoGo (2026) — Open-source project rebuilding AlphaGo from scratch using modern AI tooling; a pedagogical exploration of MCTS, self-play, and their relationship to LLM reasoning, documented in a tutorial and presented in a long-form public technical discussion.

Awards & Recognition

  • ICLR 2017 acceptance — Gumbel-Softmax paper accepted as a poster at one of the field’s most selective venues.
  • WIRED feature — Work on opening latched doors with neural networks was featured in WIRED magazine.
  • Google AI Blog features — Multiple projects (Semantic Grasping, QT-Opt, Grasp2Vec, RetinaGAN, BC-Z) were featured on the Google AI blog, an internal recognition of high-impact research.
  • Area Chair — Recognised by ICML, CoRL, and NeurIPS as an area chair.

Key Relationships

  • Shixiang Gu & Ben Poole — Co-authors of the Gumbel-Softmax paper; Gu introduced the Gumbel-Max trick, Poole contributed the semi-supervised experiments and mathematical derivations.
  • Sergey Levine — Frequent collaborator at Google on QT-Opt, Grasp2Vec, and BC-Z; a leading figure in robot learning.
  • Vincent Vanhoucke — Google Brain principal who encouraged Jang to pursue the Gumbel-Softmax idea; co-author on semantic grasping and QT-Opt.
  • Chris Maddison, Andriy Mnih & Yee Whye Teh — Independent co-discoverers of the Concrete Distribution at DeepMind; the simultaneous discovery is a notable episode in ML history.
  • Berndt Børnich — CEO and founder of 1X Technologies; Jang joined Børnich’s company as its first California employee and helped execute the company’s pivot to home humanoids.
  • Dwarkesh Patel — Hosted a long-form podcast with Jang in May 2026 covering AlphaGo, MCTS, and RL in LLMs; the conversation also reflected on the relationship between search and learning more broadly.

Personal Style

Jang writes with unusual directness about the practical realities of research and career decisions — his blog posts on leaving Google Brain and leaving 1X are candid first-person accounts of the reasoning and uncertainty involved, a rare register in a field dominated by announcement-style communication. His technical sensibility is grounded in a conviction he calls the “Bitter Lesson” perspective: rich data combined with simple scalable algorithms consistently outperforms careful engineering, and identifying the next “magical object” — a new model family or capability with surprising generalisation — is the highest-leverage act in AI research. He is a prolific technical blogger with a distinctive voice that mixes systems-level engineering detail with broader philosophical speculation about intelligence, simulation, and computation. His public following reflects an audience that spans practitioners seeking tutorials and researchers interested in his perspective on the direction of embodied AI.

References