Weiji Xie

Introducing our recent work, TextOp: Real-time Interactive Text-Driven Humanoid Robot Motion Generation and Control TL;DR * Control humanoid robots via real-time revisable text prompt * Seamless switching between multiple skills Paper: text-op.github.io/static/pdf/pap… Check website&code

@ChongZitaZhang the actuator network looks like scaffolding: people add it when they haven't figured out true the sim‑to‑real gap, and take it out once they have.

We need actuator network -> we don't need actuator network -> we need actuator network...

@MozarellaPesto @lucasmaes_ Interesting work! My question is how can you learn the action decoder without the action label? do you add another supervised learning loss for it?

I extended LeWorldModel by @lucasmaes_ et al to no longer require action conditioning. It now learns controllable dynamics directly from raw video. It uses latent action modelling to self-learn discrete control codes between latent states. Hence 'La leWorldmodel' 👇👇👇

@JonEmbleyRiches Cool! just wondering, is there any plan to support parallel simulation & RL training?

Releasing the alpha of Unreal Robotics Lab — an open-source Unreal Engine plugin with full MuJoCo physics. Photorealistic rendering and accurate contact physics. No compromises on either side. GitHub: github.com/URLab-Sim/Unre… Paper: arxiv.org/abs/2504.14135

@lok_s_h @jinrui82 thx for your like! seems you rewrite the original c++ deploy code in python, welcome to PR to TextOp repo if you like :)

kudos to @Shi_Soul , @jinrui82 and the team for TextOp.

agentic engineering hardware code is thrilling. zero-shot port (CRC checks, FSM, motion loader, ob manager, etc). Prompt Jitter Prompt Done

@techvibekorea thx for your interest :)

@Shi_Soul Mind blown! Text-driven skill switching for humanoids just gave the creative runway a tech upgrade. I can totally picture this debuting live onstage or elevating next-gen fashion shows.

Introducing our recent work, TextOp: Real-time Interactive Text-Driven Humanoid Robot Motion Generation and Control TL;DR * Control humanoid robots via real-time revisable text prompt * Seamless switching between multiple skills Paper: text-op.github.io/static/pdf/pap… Check website&code

Project Page: text-op.github.io

Open Source Code: github.com/TeleHuman/Text…

Introducing a preview version of TextOp: Real-time Interactive Text-Driven Humanoid Robot Motion Generation and Control Project Website: text-op.github.io Open Source Code: github.com/TeleHuman/Text…

@parada_car88104 @GoogleDeepMind amazing!

📣 #MuJoCo Announcement 📣 : MuJoCo Warp is now available in MJX for all MuJoCo Playground environments, achieving up to 2x speedup in more complex contact-rich scenes. Check it out and have fun! @GoogleDeepMind Robotics github.com/google-deepmin…

@maxleedev amazing! looking forward to this feature for long

current chat interfaces suck, so i built a canvas for llms

🔥 Introducing KungfuBot – our whole body control framework for humanoid robots, now released & open-sourced! It enables high-speed, dynamic martial arts-like motions (e.g., spinning kicks) on platforms like Unitree G1. 🏠 kungfu-bot.github.io #OpenSource #HumanoidRobot #AI

@ZhongyuLi4 seems the feasible motions are too few( walk, run, wave) can we do sth more?

Command humanoids *directly* with natural language? Introducing LangWBC, a generative, end-to-end policy that turns natural language into real-world whole-body humanoid control! 💬→🦿Smooth, robust, surprisingly intuitive! See more 👉 LangWBC.github.io #RSS2025

@pcastr @johanobandoc Interesting!

How do you get the most out of your parameters in deep RL? 🤔 @johanobandoc Aaron Courville and i say: ✂️A Pruned Network is a Good Network!✂️ arxiv.org/abs/2402.12479 We demonstrate that gradual magnitude pruning yields *dramatic* improvements in RL agents! 1/12

🥽 Want to use your new Apple Vision Pro to control your robot? Want to record how you navigate / manipulate the world to train a policy? I developed an app for VisionOS that can stream your head / wrist / finger movements over WiFi, which you can subscribe on any machines using a simple python library. 🧵[1/6]

Zero Bubble Pipeline Parallelism paper page: huggingface.co/papers/2401.10… Pipeline parallelism is one of the key components for large-scale distributed training, yet its efficiency suffers from pipeline bubbles which were deemed inevitable. In this work, we introduce a scheduling strategy that, to our knowledge, is the first to successfully achieve zero pipeline bubbles under synchronous training semantics. The key idea behind this improvement is to split the backward computation into two parts, one that computes gradient for the input and another that computes for the parameters. Based on this idea, we handcraft novel pipeline schedules that significantly outperform the baseline methods. We further develop an algorithm that automatically finds an optimal schedule based on specific model configuration and memory limit. Additionally, to truly achieve zero bubble, we introduce a novel technique to bypass synchronizations during the optimizer step. Experimental evaluations show that our method outperforms the 1F1B schedule up to 23% in throughput under a similar memory limit. This number can be further pushed to 31% when the memory constraint is relaxed. We believe our results mark a major step forward in harnessing the true potential of pipeline parallelism.