Yishu Li ✈️ CoRL 2026 (@LisaYishu) header
Yishu Li ✈️ CoRL 2026 (@LisaYishu) profile picture

CS PhD @Berkeley_ai | Prev MS @CMU_Robotics, CS Undergrad @Tsinghua_Uni

Berkeley, CA
Joined October 2015
Photo by Yishu Li ✈️ CoRL 2026 (@LisaYishu)Photo by Yishu Li ✈️ CoRL 2026 (@LisaYishu)Photo by Yishu Li ✈️ CoRL 2026 (@LisaYishu)Photo by Yishu Li ✈️ CoRL 2026 (@LisaYishu)Photo by Yishu Li ✈️ CoRL 2026 (@LisaYishu)Photo by Yishu Li ✈️ CoRL 2026 (@LisaYishu)Photo by Yishu Li ✈️ CoRL 2026 (@LisaYishu)Photo by Yishu Li ✈️ CoRL 2026 (@LisaYishu)Photo by Yishu Li ✈️ CoRL 2026 (@LisaYishu)Photo by Yishu Li ✈️ CoRL 2026 (@LisaYishu)
Robots can’t always see the physics they act under: hidden mass, friction, kinematic constraints. A scalar reward only says "failed." The outcome of an action says much more. Introducing SCOUT: test-time policy adaptation from action-outcome feedback. Sim-trained, adapts in the real world in a few tries 🧵👇 #CORL2026
6
26
3
127
9,891
I'm at #CoRL2026 this week and excited to chat about test-time adaptation and continual learning! I’ll be presenting our work SCOUT at: 📍 Poster 110 on Wednesday from 10:30-12 at the JW Marriott (3rd floor)!! 📍 Poster at Pretrain to Adapt Workshop 3:30-4:15 at JW Marriott (301-302)! Come chat! Happy to meet old and new friends :D
Robots can’t always see the physics they act under: hidden mass, friction, kinematic constraints. A scalar reward only says "failed." The outcome of an action says much more. Introducing SCOUT: test-time policy adaptation from action-outcome feedback. Sim-trained, adapts in the real world in a few tries 🧵👇 #CORL2026
1
1
34
865
Yishu Li ✈️ CoRL 2026 retweeted
🚨Introducing DISCO🪩: Distilled Tree Search over Continuous Actions (Best Paper Finalist🏆 at #CoRL2026) DISCO combines learned policies with Monte-Carlo Tree Search (MCTS) over continuous actions for long-horizon manipulation tasks. A thread 🧵👇
6
29
3
140
8,519
Yishu Li ✈️ CoRL 2026 retweeted
1/ Flow matching has been very successful for robotics and RL policies. Can we use the same recipe for value functions? Prior methods have tried, but they unfortunately produce biased, collapsed predictions. We find that the key to accurate value estimation is to enforce Bellman consistency at every noise level. Introducing Value Forcing, an accurate value function estimator even with very small amounts of data. 🧵
6
59
4
435
43,702
Yishu Li ✈️ CoRL 2026 retweeted
Introducing DreamFAR, a world model trained from scratch using only robot data, without internet pretraining. It supports: 🎮 Real-time teleoperation ⏱️ Minutes of long-horizon rollout 🧩 Generalization across 1,000+ tasks yimingdou.com/DreamFAR 🧵 (1/8)
24
54
21
306
33,295
Yishu Li ✈️ CoRL 2026 retweeted
(𝐵𝑎𝑐𝑘𝑔𝑟𝑜𝑢𝑛𝑑 𝑚𝑢𝑠𝑖𝑐: 𝑎 𝑙𝑖𝑡𝑡𝑙𝑒 𝑑𝑢𝑒𝑡 𝑝𝑙𝑎𝑦𝑒𝑑 𝑏𝑦 𝑚𝑒 𝑎𝑛𝑑 𝑡ℎ𝑒 𝑟𝑜𝑏𝑜𝑡 𝑖𝑡𝑠𝑒𝑙𝑓 🎶 𝐸𝑛𝑗𝑜𝑦 𝑡ℎ𝑒 𝑣𝑖𝑑𝑒𝑜 𝑤𝑖𝑡ℎ 𝑠𝑜𝑢𝑛𝑑 𝑜𝑛! 🔊) Ever dreamed of a robot that could play the harp 🎼, tie your shoelaces 👟, or even perform delicate wet-lab experiments 🧪? The key is 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐧𝐨𝐭 𝐣𝐮𝐬𝐭 𝐡𝐨𝐰 𝐡𝐮𝐦𝐚𝐧𝐬 𝐦𝐨𝐯𝐞, 𝐛𝐮𝐭 𝐡𝐨𝐰 𝐭𝐡𝐞𝐲 𝐟𝐞𝐞𝐥 𝐚𝐧𝐝 𝐜𝐨𝐧𝐭𝐫𝐨𝐥 𝐟𝐨𝐫𝐜𝐞𝐬. We introduce 𝐅𝐀𝐂𝐓, a Force–Action–Track foundation model that learns from 2,000 hours of force-rich human demonstrations in our EgoForce dataset, bringing us closer to robots with truly human-like dexterity! 🩷💙💜 arxiv.org/abs/2610.19155 Bigggggg shoutout to @ZhiCao01756604 !
4
31
8
151
11,129
Yishu Li ✈️ CoRL 2026 retweeted
🌍 Excited to share T2-z: Test-Time Optimization of Environment Latents for Robotic World Models! Same observation, same action, different environments → different outcomes. How can world models adapt to hidden environmental factors like friction, mass, or physical rules? 💡 T2-z conditions world models on optimizable environment latents. T2-z directly optimizes the conditioning latents through world prediction, without supervision of what these factors are, what values they take, or how many factors exist. 📊 Meaningful latent structure emerges purely through world prediction! For compounded environmental factors (mass × friction), PCA of the learned latents reveals a two-dimensional structure that closely reflects the underlying factor distribution, despite the model never knowing how many factors are present or what they represent. 🤖 Across 7 simulated settings and 4 real-world robotic tasks, T2-z: •Reduces object prediction error from 15.20 to 7.00 (54% reduction). •Improves action success from 33.73% to 72.96% (+39.23 percentage points). 🌐 Project: t2-z-latent-optimization.git… 📄 Paper: arxiv.org/abs/2610.22364 💻 Code: github.com/T2-z-latent-optim… Collab with Chenyuan, @RavenHuang4 @YunfanJiang @wenlong_huang @davheld @vitorguizilini @yuewang314 @drfeifei @jiajunwu_cs , leo and @YejinChoinka
8
32
5
164
19,005
Yishu Li ✈️ CoRL 2026 retweeted
Can we make dexterous robots compliant without sacrificing precise control? Excited to introduce FOMO-Dex: Force-Modulated Dexterity! 🖐️ FOMO-Dex brings force-aware compliance to dexterous teleoperation, enabling robots to pick up thin cards and chips 🃏, cut paper 🔪, turn nuts 🔩, and use screwdrivers 🪛. These force-aware demonstrations also train better autonomous policies! 🎥 Videos & paper: yingyuan0414.github.io/fomo-… 🧵👇 (1/8)
11
41
4
168
14,070
Simulation: six tasks with hidden dynamics. SCOUT needs the fewest attempts on all tasks, beating all baselines, including naive history conditioning.
1
2
454
Yishu Li ✈️ CoRL 2026 retweeted
Last year I trained the advantage-conditioned VLA for π*0.6 during my internship @physical_int. Now, I have a much improved recipe. Introducing ✨Advantage Guidance✨, a performant RL method with stable advantage-conditioned supervised training and contrastive guidance sampling
10
66
9
609
66,571
Yishu Li ✈️ CoRL 2026 retweeted
Doing proper science in robotics is hard, because human demonstrations are expensive and less controllable. OCBench is a new benchmark providing "human-like" scripted policies. This enables scientific studies of BC and RL with virtually unlimited data!
29
68
9
605
32,705
Yishu Li ✈️ CoRL 2026 retweeted
[1/N] Large flow policies (e.g., VLAs) have become extremely popular, unlocking scalability. However, they come at a cost: the robot cannot react quickly. π𝐑² enables them to achieve reactive, real-time control.
10
38
3
137
26,942
Yishu Li ✈️ CoRL 2026 retweeted
#ICML2026Spotlight Standard diffusion is powerful, but inference latency is a real headache for real-time robot control. HDFlow uses rectified flow for instant trajectory generation in long-horizon tasks. Fantastic work by @girish_432 @AveryJuuu0213 !
🎉@icmlconf 2026 Spotlight! 🤖 Can generative planners be highly exploratory for long-horizon tasks & fast for real-time control? Single-paradigm models struggle with speed. Introducing HDFlow: Hierarchical Diffusion-Flow Planning! We combine the best of both worlds. (1/7)
4
17
2,378
Yishu Li ✈️ CoRL 2026 retweeted
Boo! 👻 GHOST has been accepted to #RSS2026! Can we learn manipulation skills from human video (without action retargeting)? Yes! GHOST learns generalizable manipulation skills by training a hierarchical policy with an embodiment-agnostic sub-goal predictor + embodiment-specific controller. 🧵(1/7)
2
14
1
71
4,665
Yishu Li ✈️ CoRL 2026 retweeted
🧩 #ICML2026 💥 How can a model discover the 3D objects in a scene—their shape, color, and position—without any labels? Introducing 3D-DLP, a self-supervised object-centric model that decomposes colored 3D scenes (RGB-D and voxels) into a set of 3D latent particles.
4
9
2
54
11,120