Yufeng Tian
20 posts
Yufeng Tian
@TianRainWind
Incoming PhD of MMLab@CUHK.
Katılım Ağustos 2022
127 Takip Edilen42 Takipçiler
🦾 Introducing ZPRL: RL steers pretrained manipulation policies through a compact bottleneck latent, with smooth and efficienct exploration.
More can be found in:
Website: manutdmoon.github.io/ZPRL/
Code: github.com/manutdmoon/ZPRL
Paper: arxiv.org/abs/2605.19919
English
Eyes see. Hands feel. Many visuotactile methods rely only on concatenation, leaving room to better exploit alignment and complementarity for manipulation.
Excited to share our #ICRA2026 paper: "ViTaS: Visual Tactile Soft Fusion Contrastive Learning for Visuomotor Learning".
English
@shuiqi28 @HarryXu12 @fancy_yzc @still_wtm @Will_Zhou2025 @Yuanhang__Zhang @ZixianLiu03 6/N
Paper and code have been released. Check them out!
Paper: arxiv.org/pdf/2602.11643
Code: github.com/SkyRainWind/Vi…
English
5/N
ViTaS is co-led by Shuiqi Cheng @shuiqi28, advised by Huazhe Xu @HarryXu12 and Zhecheng Yuan @fancy_yzc. Also shoutout to Tianming Wei @still_wtm , Tianxing Zhou @Will_Zhou2025, Yuanhang Zhang @Yuanhang__Zhang, Zixian Liu @ZixianLiu03 and Qianwei Han for their great effort!
English
Results: For RL, ViTaS shows significant improvement in 12 tasks. For IL, ViTaS still wins across 3 simulation IL tasks, showing that touch compensates for reduced vision.
English
2/N ViTaS in RL and IL.
In RL we use ViTaS as the feature extractor with PPO. In IL we plug ViTaS into DP, replacing its default encoders with our visuo-tactile pipeline.
Fair comparison: DP gets 3 cameras, while ViTaS uses only 1 head camera + tactile sensors.
English
Yufeng Tian retweetledi
Real-world deployment is crucial for robotics.
I wrote about a recipe: strong base model + RL post-training (offline & online) → toward human-level capability.
lei-kun.github.io/blogs/RL100.ht…
English
Achieving 100% success rate is really brilliant!
Kun Lei@kunlei15
Introducing RL-100: Performant Robotic Manipulation with Real-World Reinforcement Learning. lei-kun.github.io/RL-100/ 7 real robot tasks, 900/900 successes. Up to 250 consecutive trials in one task, running 2 hours nonstop without failure. High success rate against physical disturbances, zero-shot, and few-shot adaptation Our first step toward a deployable robot learning system.
English
@KYamiuu49 @Kemo_monkuri Best illustration on ao-chan I've ever seen.
English
【お知らせ】
┈┈┈┈┈┈┈┈┈┈┈┈
劇場版『#クドわふたー』
3月より見放題配信決定！
┈┈┈┈┈┈┈┈┈┈┈┈
各動画配信サービスにて3月1日より順次配信開始！
kud-wafter.jp
日本語