Roger Qiu (@RogerQiu_42) profile picture

Researcher at Embodied Superintelligence at Meta. Previously CS @ Illinois and ECE @ UCSD. https://teapawt.net/t.co/aWIV9ZvbHT

San Diego, CA
Joined March 2024
Photo by Roger Qiu (@RogerQiu_42)Photo by Roger Qiu (@RogerQiu_42)Photo by Roger Qiu (@RogerQiu_42)Photo by Roger Qiu (@RogerQiu_42)Photo by Roger Qiu (@RogerQiu_42)Photo by Roger Qiu (@RogerQiu_42)Photo by Roger Qiu (@RogerQiu_42)Photo by Roger Qiu (@RogerQiu_42)Photo by Roger Qiu (@RogerQiu_42)Photo by Roger Qiu (@RogerQiu_42)
Diverse training data leads to a more robust humanoid manipulation policy, but collecting robot demonstrations is slow. Introducing our latest work, Humanoid Policy ~ Human Policy. We advocate human data as a scalable data source for co-training egocentric manipulation policy.⬇️
8
54
19
245
87,125
Mazeyu is one of the first few folks who I worked with at UCSD. Excited to see what they can build next!
Three months ago, I wouldn’t have imagined we could build all of this so quickly. Seeing the infrastructure, research agents, and robot learning framework come together—and start producing real results—has been incredibly rewarding. What excites me most is where this could lead. As agents take on more of the research loop, we can explore more ideas, learn from more failures, and spend more time asking the questions that matter. Each experiment can leave us with better models and a better system for doing the next one. Proud of what this team has built. There’s still a lot to figure out, but these three months have made me much more optimistic about what we can build next.
3
466
Early glimpses of agentic manipulation capability from Meta's Muse Spark 1.2
1/ muse spark 1.2 is a very strong multimodal model—it can do visual coding, robotics planning, and audio-visual understanding that all come together through agentic tools.
6
761
After 2026, “Foundation” manipulation models that don’t work out of the box on simple tasks ~= bad models
MolmoAct2 kinda just works out of the box in an unseen environment and with unseen objects. I find it really crazy that I can just download a model from the internet and plug it into my robot and it can do some tasks at a limited capacity right away. this video is at 2x speed
7
901
Join us at egoverse.ai to push the frontier of learning from egocentric data
Introducing EgoVerse: an ecosystem for robot learning from egocentric human data. Built and tested by 4 research labs + 3 industry partners, EgoVerse enables both science and scaling 1300+ hrs, 240 scenes, 2000+ tasks, and growing Dataset design, findings, and ecosystem 🧵
1
1
12
1,416
Keypoints and MANO are not the ultimate answer to dex hand manipulation. It should be a latent. The work is in modeling the latent.
Ever want to have a single policy to control diverse robots as well as different dexterous hands, or to observe the emergent behavior under cross embodiment training? Introducing our #CVPR2026 paper XL-VLA, Cross-Hand Latent Representation for Vision-Language-Action Models.
1
16
2,086
Humanoids need to mimic humans, and we made a human behavior model for it. Was fun working with @XiongyiCai
A large human behavior model. Introducing In-N-On, our latest findings in scaling egocentric data for humanoids. 1. Pre-training and post-training with human data 2. 1,000+ hours of in-the-wild data and 20+ hours of on-task data with accurate action labels Website: xiongyicai.github.io/In-N-On… Arxiv: arxiv.org/abs/2511.15704 By simply scaling data, our robot can follow novel language instruction. Check out the 🧵
4
16
3,118
Compliance is a must for robot working in a home setting.
What if a humanoid robot could choose how to interact with the environment 🤖 — soft when it needs compliance, stiff when it needs precision, and force-aware when it must push/pull? That’s exactly what our Heterogeneous Meta-Control (HMC) framework enables. Our new framework learns to blend and route across position, impedance, and hybrid control from real demonstrations, letting a humanoid wipe, push, lift, and insert with smoothness and stability in contact-rich & forceful scenarios. Website: 👉 loco-hmc.github.io/
5
524
Excited to see Deformable Real2Sim in action!
🧵 Evaluating robot policies in the real world is slow, expensive, and hard to scale. During my internship at @SceniXai this summer, we had many discussions around the two key questions: how accurate must a simulator be for evaluation to be meaningful, and how do we get there? Our new framework, Real2Sim-Eval, takes a step toward that answer. By combining Gaussian Splatting for photorealistic rendering and soft-body digital twins for realistic dynamics, we make simulation predictive of real-world performance. 👉 real2sim-eval.github.io
1
8
1,358
Sim-real visual difference is closing fast
Ever want to enjoy all the privileged information in sim while seamlessly transferring to the real world? How can we correct policy mistakes after deployment? 👉Introducing GSWorld, a real2sim2real photo-realistic simulator with interaction physics with fully open-sourced code.
2
5
1,573
Robots will always work with near-perfect depth in 3 years.
🚀 Want to build a 3D-aware manipulation policy, but troubled by the noisy depth perception? Want to train your manipulation policy in simulation, but tired of bridging the sim2real gap by degenerating geometric perception, like adding noise? Now these notorious problems are gone with our Camera depth Models! The Camera Depth Models (CDMs) can be plug-in modules in a real robot pipeline, transforming noisy depth into high-quality perception, enabling seamless sim-to-real transfer, making real robot manipulation work as in simulation! 🎯 Why it matters: Accurate geometry with CDMs helps a sim-data-driven policy solve a set of complex, long-horizon tasks from 0% to 85%+ success! Now you can even train in simulation, deploy on real robots WITHOUT further domain adaptation. Just plug in our CDMs to your existing pipeline! ✨ Highlight: • Zero-shot sim-to-real transfer with 73%+ success (vs 0% baseline) • Depth-only imitation learning achieves 85%+ success • Works with RealSense D435/L515, Kinect, ZED2i & more 🛠️ Everything is open: • We open-source CDMs for 5 distinct cameras • We open-source the collected ByteCamDepth Dataset, which contains 170K+ RGB-depth pairs across 7 cameras & 10 configurations, a comprehensive real-world depth dataset. • We open-source our codes for sim-to-real, camera depth model inference. We also share our modular, real-robot control framework designed for robotic manipulation, which provides you a unified interface for controlling various robot arms, integrating sensors, and executing policies in real-time! • We also made a clean sim-to-real tutorial based on our framework! Check everything and interactive demos in manipulation-as-in-simulatio… We expect CDMs to be a foundation of your daily robotic research! #Robotics #ComputerVision #SimToReal #DepthPerception #OpenSource
1
4
883
Genie 3 shows amazing 3D consistency even across multiple prompts. You can step inside a garden. Merge videos generated by different text prompts into the same world, and it just works. It gives you RGB renderings and consistent 3D features. Done in 30mins⬇️
2
8
70
6,817
1. Gemini CLI breaks down the videos into images. 2. Merge all images into the same colmap runs 3. Run Feature Splatting (feature-splatting.github.io/) That's it!
1
8
335
Join our CoRL workshop in Seoul to check out the latest progress in learning from humans - the next potentially scalable data source for robot learning!
Current AI models only learn from a fraction of human intelligence. At CoRL 2025, our brand new "Human to Robot (H2R)" workshop explores how robots can learn from the vast, untapped physical human experience. sites.google.com/view/h2r-co… Extended abstract / paper submission deadline Aug 15th. Co-organized with @xiaolonw, @simar_kareer, @RogerQiu_42, Sha Yi, James Fort, @NimaFazeli7, and Jianlong Ye
4
337
For robotics, business applications like manufacturing will come before home applications. Co-designing hardware and software may be one way to go and under-estimated rn
For years, I’ve been tuning parameters for robot designs and controllers on specific tasks. Now we can automate this on dataset-scale. Introducing Co-Design of Soft Gripper with Neural Physics - a soft gripper trained in simulation to deform while handling load.
2
390
Chatting with @chris_j_paxton and @micoolcho on learning from human. Quite aligned to Tesla’s recent progress.
Timely episode given the recent progress updates from Tesla on using human data (without teleop) to train Optimus. Tks @RogerQiu_42 for coming on the pod with me & @chris_j_paxton; really cool work!
1
4
678
Check out our cool humanoid robot that reaches for objects on the ground!
Meet 𝐀𝐌𝐎 — our universal whole‑body controller that unleashes the 𝐟𝐮𝐥𝐥  kinematic workspace of humanoid robots to the physical world. AMO is a single policy trained with RL + Hybrid Mocap & Trajectory‑Opt. Accepted to #RSS2025. Try our open models & more 👉 amo-humanoid.github.io
11
837
We show spatial memory with an attention mechanism as an effective alternative to simple feature distillation.
[1/n] We are releasing M3 (#ICLR2025): a Gaussian Splatting method that builds LMM memories for arbitrary scenes. 🔥 [Efficient] 16 degrees in each Gaussian primitive for one LMM. 🔥 [Alignment] The rendered features are directly in the source LMM embedding space.
1
10
1,125
Feature Splatting can now turn Objaverse assets into GS. With optimized kernel, 30K iters can be done in <1min on a single 4090 GPU.
3D Gaussians Splats look good, but they are less interesting when static. Our ECCV 2024 work, Feature Splatting, enhances 3DGS with foundational vision features and physical simulation to synthesize dynamic scenes from static 3D captures. ✨Open-sourced in NerfStudio!👇
11
1
45
8,672
Spending a couple of hours to make a model run faster always feels so good 🧐
6
413
Check out how smooth it is
An interesting human tele-operated robot-robot interaction demo w/ @xuxin_cheng, @Jialong_LI_UIM, @AaronYANG2000
1
4
2,708