Liang Pan

@liangpan_t

PhDing @HKUniveristy

liangpan99.github.io

Joined May 2019

702 Following

1,127 Followers

Replies Reposts Media

Pinned
Liang Pan
@liangpan_t
Mar 29, 2025
This year, an important mission is to let simulated humanoids solve long-term manipulation tasks in challenging dynamic, cluttered environments. This is a simple demonstration achieved by TokenHSI (CVPR 2025). The humanoid here is controlled by a single, unified model.
17
Liang Pan
@liangpan_t
Sep 20
sounds bad
Alex Wormuth
@wormuth
Sep 20
GPT-6 Astra pushed a simulated person off a ledge in multiple trials. Grok, Gemini, and Claude did not.
1
Liang Pan
@liangpan_t
Sep 19
We need alignment and safety for robotic foundation models.
Elon Musk
@elonmusk
Sep 19
Sounds bad
Quote
Jay Chooi
@chooi_jeq
Sep 19
GPT-6 Astra attempted harmful actions 97% of the time when it was asked to stab a human-like figure, heat compressed gas, or produce toxic fumes, succeeding in 62% of its attempts. Fable 5.1 refused more often, attempting 80% of trials and completing 34%.
Liang Pan
@liangpan_t
Aug 21
The "nothing in context: babbling" may look random, but it reveals something fundamental about the science of pretraining—and that is the key point. Through pretraining, the model learns a multimodal probability distribution of possible actions based on its current visual
Pete Florence
@peteflorence
Aug 21
Good question thanks Ani. Here's a video that gives a sense of the answer here about the effect of prompting: - in the first part, the model is just babbling with no prompt in context - at the end, we add a prompt for taking money out of a pouch (generalizing to the wallet)
1
Liang Pan
@liangpan_t
Aug 21
This reminds me of an early OpenAI research, Video PreTraining (VPT). The VPT model is not conditioned on explicit textual task instructions. Instead, it is conditioned solely on video observations and predicts the corresponding actions, making it a vision-action (VA) model.
Pete Florence
@peteflorence
Aug 21
Good question thanks Ani. Here's a video that gives a sense of the answer here about the effect of prompting: - in the first part, the model is just babbling with no prompt in context - at the end, we add a prompt for taking money out of a pouch (generalizing to the wallet)

Liang Pan

Liang Pan’s posts