ControlWiz

136 posts

ControlWiz banner
ControlWiz

ControlWiz

@Control_wiz

Building & thinking out loud

شامل ہوئے Nisan 2026
32 فالونگ8.3K فالوورز
ControlWiz
ControlWiz@Control_wiz·
Useful specialized robots already exist in many homes. A robot that can do everything a human can do is likely not happening in our lifetime. The goal isn’t to replicate humans. It’s to build robots that help humans solve problems. Legged humanoids that look like humans are not the best solution for many applications.
RoboStrategy@RoboStrategy

How soon until humanoid robots are part of everyday life?    CEO Andrew Kang (@Rewkang) gives his thesis:

English
1
1
6
994
ControlWiz ری ٹویٹ کیا
Junyao Shi
Junyao Shi@JunyaoShi·
Basically you pick your poison in terms of where you want to inject human inductive bias in the loop: directly at low level (imitation) or at some abstraction level (reward shaping, initialization etc.). There are arguments for either to be more bitter lesson pilled. One bitter lesson for RL in robotics manipulation is: it is actually often less scalable due to the amount of task-specific and environment specific tuning of each parameter to get anything working.
Yunlong Song@realyunlong

Most robotics RL paper is often just imitation learning in disguise. The "human expert" transfer task through extensive reward shaping, curricula, initialization strategies, environment design, and various tricks. You are providing demonstrations--just indirectly. A reward function is just a demonstration written in a different language.

English
1
2
13
2.9K
ControlWiz
ControlWiz@Control_wiz·
@yacineMTB This would be an issue for open loop unstable systems or if you have fast disturbances/uncertainties.
English
1
0
5
1.1K
kache
kache@yacineMTB·
I wonder if anyone has ever let a policy decide how much time it wants to wait before making its next action
English
21
3
119
10.5K
ControlWiz
ControlWiz@Control_wiz·
@lymytom20 I haven't followed more recent work, but Drone racing and RL locomotion have been established since ~2018. Locomotion on relatively structured terrain is a fairly mature problem, and MPC variants have also shown strong results. Backflips are cool demos, but nothing more.
English
1
0
1
38
Tommy Ly
Tommy Ly@lymytom20·
@Control_wiz What's your take on all the recent papers using RL for locomotion - from drone races to Unitree backflip?
English
1
0
0
33
ControlWiz
ControlWiz@Control_wiz·
Always do MPC first. A lot of useful problem can be mapped to an MPC. RL is inefficient.
English
3
0
21
1.7K
ControlWiz ری ٹویٹ کیا
bradford
bradford@LusciousPear·
Pixels are not enough. Robots need a realtime learning loop too, but it needs to use kinematic and sensor data from the real world. since that's where robots actually operate.
English
0
2
13
861
ControlWiz ری ٹویٹ کیا
inControl podcast
inControl podcast@inControlpdcst·
New episode! 🎙️Peter Caines (@mcgillu) takes us through six decades of control, from system identification to adaptive control to the birth of mean field games, all infused with the restless spirit of the 1960s: Coltrane, Kerouac, and even hitch-hiking all the way to Istanbul!👇
inControl podcast tweet media
English
0
3
7
398
ControlWiz ری ٹویٹ کیا
Naval
Naval@naval·
Science is not a process, a credential, or an institution. It is the unflinching pursuit of truth, carried out by the few, co-opted by the many.
English
766
2.3K
14.5K
1M
ControlWiz ری ٹویٹ کیا
Baxate
Baxate@Baxate·
friendly reminder that things get way less scary when you understand how to use them and even less so when you understand how they work and it serves the dual purpose of making you more resilient to people manipulating that fear to progress their own agenda
English
3
5
53
2.8K
ControlWiz
ControlWiz@Control_wiz·
AGI doesn’t have a set definition, let alone be anywhere on the horizon. Treating it as an inevitable trillion dollar superweapon is science fiction and only adds fear and confusion.
gabriel@gabriel1

agi is the most economically valuable asset of all time, there will be trillions in free market capital put into it this is extremely unlike the manhattan project. this time, governments can only cooperate. we can't just pick a winner, or that winner will lose

English
0
0
4
1.1K
ControlWiz ری ٹویٹ کیا
Mathieu
Mathieu@miniapeur·
Being a PhD student is hard for many reasons. On paper, the flexibility can seem like an advantage: you often do not have a fixed schedule or many strict deadlines. In practice, however, that same freedom can become a trap. Since your progress depends so much on your own initiative, it is easy to convince yourself that every extra hour of work is an investment in your future. For students who are already inclined to overwork, this can blur the line between dedication and self-exploitation, eventually leading to burnout.
English
8
26
381
27.6K
ControlWiz
ControlWiz@Control_wiz·
VLA based models can only interpolate over the support of its training/pre-training distribution. To become a generalist, it would need training coverage dense enough to span an enormous (possibly unbounded) task space, which is not scalable. A generalist needs to learn on the job and plan accordingly.
English
4
3
28
2.8K
Lavnir
Lavnir@Lavnir_·
@Control_wiz Just prompt the controller with be an expert driver don’t crash lol
English
1
0
0
76
ControlWiz
ControlWiz@Control_wiz·
I have to stop arguing on here.
Sparr@sparr0

@Control_wiz What work? This wouldn't even warrant a publication. The implementation would just be things like "don't ever follow closely enough that them braking could cause a collision", "don't ever enter their turning radius", etc. all the worst case bounds for any more-informed simulation

English
2
0
6
1.6K
ControlWiz ری ٹویٹ کیا
Deb Raji
Deb Raji@rajiinio·
I do not want to do AI research that is reactive to what these companies are doing, or even what they're saying. The entire field keeps chasing after product releases. Some spend more time reading marketing copy than their colleagues papers and I just... do not want to do that?
English
5
16
127
7.7K
ControlWiz
ControlWiz@Control_wiz·
@sparr0 Never heard of such work, and highly doubt it exist. Share references.
English
1
0
0
190
Sparr
Sparr@sparr0·
@Control_wiz You can predict their entire navigational envelope and avoid it completely.
English
1
0
0
143
ControlWiz
ControlWiz@Control_wiz·
In multi-agent planning you react and replan continuously using your belief states at 30–50 Hz. If the behavior models of the other agents are known, many methods exist. When they are unknown, you need some predictive model, and collision avoidance still cannot be guaranteed without strong assumptions. I’m not aware of a general method that guarantees safe behavior without these, please share if you do.
Sparr@sparr0

@Control_wiz @ChenTessler No, you don't need some assumption about how other agents will act. It is possible to drive such that you always have an escape route. To never follow so closely that you can't stop or evade if they brake. To never cross ahead of someone that might speed up.

English
1
0
5
1.8K
ControlWiz ری ٹویٹ کیا
Fei-Fei Li
Fei-Fei Li@drfeifei·
Scientific research is fundamental to advancing civilization and helping people globally to solve the most critical problems, from medicine to materials, from brain science to physics, and much beyond. This is only possible when scientists have access to the best tools of the time to conduct scientific research, including having access to AI-based tools.
English
120
471
3.1K
193.2K
ControlWiz
ControlWiz@Control_wiz·
@ChenTessler AVs also perform worse with bad weather , road debris, and mixed traffic. In normal driving conditions, they’re often better than the average human driver.
English
0
0
0
11
Chen Tessler
Chen Tessler@ChenTessler·
Isn't the fact that humans are shitty drivers that cross red lights seconds after it's switched to red, kind of the proof this isn't globally accurate? So if we agree (some) humans are shitty drivers with bad internal models and inability to predict. Then the goal is to be better than the median/mean and that's already a big improvement.
English
2
0
0
23
ControlWiz
ControlWiz@Control_wiz·
Idk if human behavior is solvable. No car will likely reach level 5 autonomy, as long as there are human drivers.
alex p.@alexpdiggs

@Control_wiz just a lil speed bump.

English
6
0
11
3.4K
ControlWiz
ControlWiz@Control_wiz·
@ChenTessler I agree. Never claimed AVs are not better than humans. My point was we don't have full autonomy because of humans on the road.
English
0
0
0
19

دریافت کریں

@JitendraMalikCV @yacineMTB @lymytom20 @mcgillu @Lavnir_ @sparr0 @elonmusk @BarackObama