Haitong Wang

69 posts

Haitong Wang

Haitong Wang

@haitong__wang

PhD Student @UofT, Robot Learning. Prev. @CarnegieMellon | @Beihang1952.

Toronto, Ontario Katılım Mayıs 2020
301 Takip Edilen45 Takipçiler
Haitong Wang retweetledi
Jiawei Yang
Jiawei Yang@JiaweiYang118·
Two months ago, I vaguely posted a number: 0.9 FID, one-step, pixel space. Now it is 0.75, and can be even lower. Many wonder how. I thought it might end as a small FID prank: simple and deliberate. It started with one question: can FID be optimized directly, and what does it reveal? Introducing FD-loss.
Jiawei Yang tweet media
English
55
155
919
210K
Haitong Wang retweetledi
Aaron Tan
Aaron Tan@aaronistan·
We are hosting the largest showing of Lume in Palo Alto Bring your own laundry Details below
English
178
122
2.6K
879.2K
Aaron Tan
Aaron Tan@aaronistan·
We’re a small team in Palo Alto with a contrarian belief. Home robotics won’t start with everything, it will start with the few chores that actually matter, in a form people already understand. This is our story @bySyncere.
English
28
34
340
49K
Haitong Wang retweetledi
Aaron Tan
Aaron Tan@aaronistan·
Design is everything at @bySyncere Excited to share our vision for the future of home robotics tomorrow.
English
15
16
160
14.8K
Aaron Tan
Aaron Tan@aaronistan·
We just did a private demo of Lume for Jeff Bezos, Will I Am, Dario Amodei, and more. The progress over the past six months has been incredible. The @bySyncere team is so excited to launch this week.
Aaron Tan tweet mediaAaron Tan tweet mediaAaron Tan tweet media
English
57
16
571
139.1K
Aaron Tan
Aaron Tan@aaronistan·
gave a preview of the new Lume design at a16z sr alumni demo day! more tomorrow @bySyncere
Aaron Tan tweet media
English
10
2
59
3.1K
Haitong Wang retweetledi
C. Zhang
C. Zhang@ChongZzZhang·
releasing AME2: Agile and Generalized Legged Locomotion via Attention-Based Neural Map Encoding arxiv.org/abs/2601.08485 In this work, we discuss how to achieve a combination of generalization and agility in legged locomotion, and propose a general solution.
English
10
39
287
83.2K
Haitong Wang
Haitong Wang@haitong__wang·
@ChongZitaZhang yeah, I agree that with full privileged obs, active perception could emerge. I was talking about when the hidden ground-truth can't be inferred from simplified privileged obs. Critic would still be fine to learn, but for actor, variance would be too high, and not ideal for learn
English
1
0
0
40
C. Zhang
C. Zhang@ChongZzZhang·
@haitong__wang sites.google.com/leggedrobotics… I once made a quadruped do fully blind navigation on rough terrains. I just gave full states to the critic, and proprioception to the actor. The active perception just blatantly emerged from RL
English
2
0
1
221
Haitong Wang
Haitong Wang@haitong__wang·
@ChongZitaZhang Unless some obs engineering is done for the critic obs to include camera direction, FOV, etc.
English
1
0
0
92
Haitong Wang
Haitong Wang@haitong__wang·
@ChongZitaZhang it would also be an issue for critic? for example, if the camera (actor obs) doesn't see the object of interest, but the critic (privileged object states) sees that the object is close to the arm EE, it won't be able to learn the right value for pick and place. Unless some obs
English
1
0
0
94
Haitong Wang
Haitong Wang@haitong__wang·
@ChongZitaZhang Using privileged info might not be ideal? as the optimal behavior under partial obs is usually slightly different from the teacher policy with privileged info
English
1
0
0
100
C. Zhang
C. Zhang@ChongZzZhang·
@haitong__wang You can just train a critic with privileged info, actually easy and stable. You can also freeze the policy to initialize the critic.
English
1
0
1
122
Haitong Wang retweetledi
Tairan He
Tairan He@TairanHe99·
Zero teleoperation. Zero real-world data. ➔ Autonomous humanoid loco-manipulation in reality. Introducing VIRAL: Visual Sim-to-Real at Scale. We achieved 54 autonomous cycles (walk, stand, place, pick, turn) using a simple recipe: 1. RL 2. Simulation 3. GPUs Website: viral-humanoid.github.io Arxiv: arxiv.org/abs/2511.15200 Deep dive with me: 🧵
English
19
160
784
258K