Alejandro Escontrela

This account is verified. Learn more

@alescontrela

Senior RS at @GoogleDeepMind. World Models

Prev PhD @ UC Berkeley, Amazon FAR, Google Brain

San Francisco, CA escontrela.me Joined January 2012

407 Following 1.5K Followers 265 Posts

Latest posts 5

Pinned post

Alejandro Escontrela @alescontrela Oct 21, 2025

Simulation drives robotics progress, but how do we close the reality gap?

Introducing GaussGym: an open-source framework for learning locomotion from pixels with ultra-fast parallelized photorealistic rendering across >4,000 iPhone, GrandTour, ARKit, and Veo scenes!

Thread 🧵

1:02

Alejandro Escontrela @alescontrela Oct 28

Carbon has been a step-change for my productivity. It handles long-running, complicated tasks with almost zero intervention. Hats off to the Gemini team!

Zachary Nado Verified account @zacharynado Oct 28

carbon

View quoted post on X

Alejandro Escontrela @alescontrela Sep 16

Egocentric human data is the key to a major breakthrough in robotics foundation models. Huge congratulations to the GSI team, especially @AdemiAdeniji, whose work is making this vision a reality

Grounded Superintelligence Verified account @GroundedSI Sep 16

Grounded API is live.

- SOTA on hand-tracking benchmarks (< 1 cm)
- SOTA on SLAM benchmarks
- In-the-wild ego data -> enriched data in minutes
- Integration with @huggingface @LeRobotHF & @rerundotio
- Built for @BitRobotNetwork RoboCap suite

Technical report & more↓

View quoted post on X

Alejandro Escontrela @alescontrela Dec 2, 2025

Introducing Holosoma (Greek for “whole body”): an easy-to-use and unified control stack that runs across humanoids (G1, T1), quadrupeds (Go2), multiple simulators (MuJoCo, MJLab, IsaacGym/Sim), any terrain, and tasks from velocity control to full-body tracking

github.com/amazon-far/hol…

0:44

Carlo Sferrazza Verified account @carlo_sferrazza Dec 1, 2025

Sim-to-real learning for humanoid robots is a full-stack problem. Today, Amazon FAR is releasing a full-stack solution: Holosoma.

To accelerate research, we are open-sourcing a complete codebase covering multiple simulation backends, training, retargeting, and real-world inference.

View quoted post on X

Alejandro Escontrela @alescontrela Oct 29, 2025

How can we standardize conditioning signals in image/video models to achieve the iterative editing & portability that Universal Scene Descriptors provide in graphics?

Introducing Neural USD: An object-centric framework for iterative editing & control

🧵

Neural USD representation. Neural USDs are composed of per-object appearance, geometry, and pose embeddings. Pre-trained image models fine-tune on this data to then allow for independent control of each attribute, including 3D pose