Junhyeong Park (@junbrro96) header
Junhyeong Park (@junbrro96) profile picture

MS Student @kaist_ai , advised by @JinwooShin0417

Seoul, Republic of Korea
Joined January 2025
Photo by Junhyeong Park (@junbrro96)Photo by Junhyeong Park (@junbrro96)Photo by Junhyeong Park (@junbrro96)
Junhyeong Park retweeted
Oral this Thursday 🎀 If you're at #CoRL2026, catch Junhyeong at the Scaling H2R workshop β€” 8:49 AM, Rooms 301–302.
2
7
461
Junhyeong Park retweeted
βœ… Thursday, Nov 12: two workshop orals. πŸ“šH2RB: Human-to-Robot Transfer via Bridging Representation and Action Space in VLAs β€” Junhyeong Park (@junbrro96) et al. Adapts pretrained VLAs from unpaired human and robot demos via geometry-anchored observation alignment and a decoupled-decoder tokenizer that shares action latents. πŸ“ Oral Β· Nov 12, 8:49–8:56 AM CST Β· JW Marriott, Rooms 301–302 πŸͺ§ Scaling H2R workshop: scaling-h2r-corl.github.io/ 🌐 junhyeong-park.me/h2rb/ πŸ“„ arxiv.org/abs/2611.04391 πŸ“šRoboCurate: Harnessing Diversity with Action-Verified Synthetic Data for Robot Learning β€” Seungku Kim (@SeungkuKim9), Suhyeok Jang (@gloryhyeok24) et al. Filters video-generated robot data by replaying each clip's predicted actions in a simulator and keeping only those whose motion matches the generated video. πŸ“ Oral Β· Nov 12, 5:00–5:45 PM CST Β· JW Marriott, Room 208 πŸͺ§ G2A workshop: actiongapworkshop.github.io/ 🌐seungkukim.github.io/robocur… πŸ“„ arxiv.org/abs/2602.18742
2
7
320
Junhyeong Park retweeted
Robots still need teleoperated demos for each new task. Human videos of the task are cheap, but visual and action gaps make them hard for robots to learn from. πŸ‘‰ RoboTryOn uses virtual try-on to turn human videos into robot demonstrations. 🌐 ryan-seungyong-lee.github.io… 🧡 [1/8]
7
10
26
1,127
Junhyeong Park retweeted
LLMs get better the longer they think. Unified multimodal models don't. Why? 🌐 rootyjeon.github.io/larc-umm… πŸ“ NeurIPS 2026 Introducing LARC: Latent Reasoning in Continuous Space. Reasoning moves into the model's hidden states, so image generation finally scales at test time. 🧡
3
27
142
23,043
Junhyeong Park retweeted
VLAs move like the teleoperators who trained them: slowly, even when they don't need to. Introducing PaceVLA πŸ€–βš‘, a VLA that picks, per action chunk, how coarsely and how far ahead to act. Up to 1.8Γ— faster, same success. πŸ“„ arxiv.org/abs/2610.25961 🌐 pacevla.github.io/ 🧡
4
23
5
110
17,956
Junhyeong Park retweeted
A recent line of WAMs generates actions from VDM hidden states of a single denoising step, yet their quality gets little attention. πŸ‘‰ We introduce FRED to improve them via self-distillation, with large gains on real robot experiments. 🌐 huiwon-jang.github.io/FRED/ 🧡 [1/7]
3
19
1
40
1,594
Junhyeong Park retweeted
Can explicitly learning future point motion improve the model that generates those future images? We introduce VPM (Video-Point Model), a 4D world model that jointly forecasts future multi-view videos + persistent 3D point trajectories. Short answer: yes. Teaching a video world model to forecast 3D point tracks makes its videos more faithful & 3D-consistent, and makes robots better. 🌐 Project: rootyjeon.github.io/video-po… 🧡 [1/6]
3
23
3
90
15,845
Junhyeong Park retweeted
πŸš€ NVIDIA Γ— KAIST Joint AI Research Lab is here! 🌟 $300M for Sovereign Agentic Future Announced by @JensenHuang himself in SF! The new lab at the Kim Jaechul Graduate School of AI brings top KAIST talent + @nvidia's full-stack tech + massive computeπŸ”₯ nvidianews.nvidia.com/news/n…
3
39
5
165
48,293
Junhyeong Park retweeted
πŸ†1st place in the AI Scientist track at Ralphthon @ ICML2026 β€” $10,000 in OpenAI credits. @happyhappy_jun & @WoominSong from @KAIST_AI We've open-sourced all of it: harness, code, the full log of research traces. See our post: byungjunyoon.ai/writing-driv… How we did it 🧡
6
18
41
5,006
Introducing TRQAM! Internalizing a KL trust region inside the sampling SDE stabilizes off-policy RL fine-tuning of pretrained flow policies. With TRQAM, we lift offline RL success on 50 OGBench tasks from 46% to 68%. 🧡 [1/8] yonghdong.github.io/blog/trq…
8
22
1
50
5,038
Junhyeong Park retweeted
RLDX-1 took the first place "officially" in the RoboCasa365 benchmark! robocasa.ai/leaderboard.html Tech report: arxiv.org/abs/2605.03269
1
29
3
58
15,908
Junhyeong Park retweeted
(1/12) Hi, we are #RLWRLD (ReaL WoRLD). RLDX-1 is live. Dexterity is intelligence, and it lives in #RLDX (RealDex). A dexterity-first foundation model from for robot hands that builds muscle memory through motion, history, and contact. - Across 10 dexterous real-world tasks on the ALLEX humanoid and DROID setup, RLDX-1 ~2Γ— outperforms Ο€β‚€.β‚… and GR00T N1.6. - SOTA on 8 simulation benchmarks in LIBERO, SIMPLER, and RoboCasa. Everything ships today: training and inference code, the pre-trained model, mid-trained checkpoints, and a fine-tuned checkpoint for every reported benchmark. LoRA recipes for parameter-efficient adaptation to your robot. Supports LeRobot v2.1 datasets. A one-line CLI to toggle motion / memory / physics modules per embodiment. Full technical report with ablations. Not just our cool demos. We document the architecture, data, and design decisions behind every result. Bring your own robot. Train and deploy RLDX-1 on it. 🧡 What's inside RLDX-1, in 12 posts. 🌐 rlwrld.ai/rldx-1 πŸ“„ arxiv.org/abs/2605.03269 πŸ’» github.com/RLWRLD/RLDX-1 πŸ€— huggingface.co/collections/R…
1
11
2
40
3,311