Ivan Kapelyukh 🇺🇦
@IvanKapelyukh
Followers
329
Following
1K
Media
6
Statuses
77
PhD student at Imperial College London, working on robot learning and computer vision. Former intern at RAI (previously Boston Dynamics AI Institute). 🇬🇧 🇺🇦
London, England
Joined July 2021
Dream2Real 🧠 is accepted at #ICRA2024! This lets robots rearrange objects "in imagination" with NeRFs and VLMs: https://t.co/dLBlHkmWXJ With: @Ren_Yifei9 @alzugarayign @Ed__Johns
3
11
72
Exciting update on Newton, a new robotics sim from Nvidia/DeepMind/Disney: would recommend following this closely!
Ssh, don't tell anyone, way too preliminary to be amplified: https://t.co/scNYdUOT5L
https://t.co/WWUVzLMsfg
1
0
15
We are extending the paper submission deadline to August 15, 2025 23:59 AOE. Submit your latest works on generalizable robot priors to the #CoRL2025 #GenPriors workshop! Webpage:
Announcing #GenPriors — the #CoRL2025 workshop on Generalizable Priors for Robot Manipulation! 📍 Seoul, Korea 📅Sat 27 Sep 2025. The workshop focuses on building generalist robot policies that work across a wide variety of environments. 🔗
0
3
7
1/ Real2Sim and Sim2Real are well-known paradigms in robotics🤖. Is there another paradigm where sim and the real world are more tightly interlinked? We introduce Real-IS-Sim: a framework featuring an always-in-the-loop, correctable simulator for behavior cloning policies.
5
48
361
MASt3R-SLAM code release! https://t.co/VFc4zyMdA3 Try it out on videos or with a live camera Work with @eric_dexheimer*, @AjdDavison (*Equal Contribution)
github.com
[CVPR 2025] MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors - rmurai0610/MASt3R-SLAM
Introducing MASt3R-SLAM, the first real-time monocular dense SLAM with MASt3R as a foundation. Easy to use like DUSt3R/MASt3R, from an uncalibrated RGB video it recovers accurate, globally consistent poses & a dense map. With @eric_dexheimer*, @AjdDavison (*Equal Contribution)
13
69
343
Thrilled to share our work @SakanaAILabs on The AI CUDA Engineer! 👷♂️ A framework that autonomously designs & optimizes CUDA kernels. Like OMNI-EPIC, we maintain an innovation archive of diverse solutions to get stepping stones that can be reused for future optimizations!
Introducing The AI CUDA Engineer: An agentic AI system that automates the production of highly optimized CUDA kernels. https://t.co/28vPHBOAX6 The AI CUDA Engineer can produce highly optimized CUDA kernels, reaching 10-100x speedup over common machine learning operations in
0
5
34
Excited to be interning at the Robotics and AI Institute (formerly Boston Dynamics AI Institute). Hoping to learn lots and work on interesting robotics research. If you're in the Cambridge/Boston area and want to meet up, let me know!
0
0
18
Instant Policy got accepted as an oral at ICLR! 🎉 Exciting times for In-Context Imitation Learning and cheap procedurally generated data! 🦾
Learning manipulation policies instantly after the demonstrations can unlock so many possibilities… 🤖 Instant Policy does just that! It efficiently tackles In-Context Imitation Learning and relies on cheap procedurally generated data! https://t.co/usWRrXME4j 1/
3
4
36
Imagine a robot that learns all sorts of tasks simply by watching you. This future is closer than ever: R+X was accepted at ICRA 2025. From an hour long POV video we can perform in-context imitation learning, from human to robot. No training or robot data needed.
5
36
243
Introducing MASt3R-SLAM, the first real-time monocular dense SLAM with MASt3R as a foundation. Easy to use like DUSt3R/MASt3R, from an uncalibrated RGB video it recovers accurate, globally consistent poses & a dense map. With @eric_dexheimer*, @AjdDavison (*Equal Contribution)
42
260
1K
The graphics community has such beautiful simulation demos. Are these running full physics like in robotics sims, or are there some physics shortcuts which still give realistic renders?
I've been doing simulation and vfx studies for years and I must say, Bifrost MPM is just mind-blowing. The simulation quality and performance are on another level. and Yes, I think it's better than Houdini. Made in @AdskMaya and Bifrost, rendered with @arnoldrenderer
0
0
7
Learning manipulation policies instantly after the demonstrations can unlock so many possibilities… 🤖 Instant Policy does just that! It efficiently tackles In-Context Imitation Learning and relies on cheap procedurally generated data! https://t.co/usWRrXME4j 1/
2
24
163
Open-source AI enabled home robot with voice chat. Tell it what to do and it will try to clean up your home for you. This is locally running Qwen2.5 on my desktop. No need to send information from your home to the cloud.
5
46
307
Very nice paper. Always exciting to see this kind of "spatial creativity" in agents!
We are very excited to announce Kinetix: an open-ended universe of physics-based tasks for RL! We use Kinetix to train a general agent on millions of randomly generated physics problems and show that this agent generalises to unseen handmade environments. 1/🧵
0
0
3
Introducing MarDini 🍸 -- our latest exploration in video diffusion models from @AIatMeta! MarDini brings an asymmetric design that breaks down video modelling into two sub-tasks: 1. A masked, auto-regressive, heavy-weight planning model focusing on long-range temporal
5
16
95
✨New #IROS2024 paper: “Adapting Skills to Novel Grasps: A Self-Supervised Approach” We leverage self-supervised data to adapt skill trajectories to novel object grasp poses. No need for depth, no prior knowledge, no CAD models, no calibration - just RGB images. Here's how👇
1
9
61
Excited to introduce Diffusion Augmented Agents (DAAGs)✨. We give an agent control of a diffusion model, so it can create its own *synthetic experience*.🪄 The result is a lifelong agent that can learn new reward detectors and policies, much more efficiently. Here's how. 👇
2
35
185
🎉 Excited to share Clarity, an open-source website template I designed to better present and visualise AI research! If you appreciate the minimalist design and want to use it in your own projects, please check it out here: https://t.co/62REQRvgs5
2
12
98
The future of robotics is in your hands. Literally. Excited to announce our new paper,✨R+X✨. A person records everyday activities while wearing a camera.A robot passively learns those skills. No labels, no training. Here's how. 👇
11
106
574
Elegant method for equivariant manipulation with impressive generalisation results!
Excited to share our new work EquiBot, which combines equivariance and diffusion policies to obtain generalizable and data-efficient policy learning. See more in the thread below for a brief explanation and links to the website, arXiv paper, and code on github.
0
0
5