Sabitlenmiş Tweet
IMLE Policy introduces a new way to train faster and more data efficient behavior cloning policies.
Will be presented at RSS2025!
imle-policy.github.io
🧵⬇️
English
Rob Lee
63 posts
Rob Lee
@roblee_rl
trying to make robots useful @sydekickbot. RL, IL etc. prev woven by toyota, everyday robots, google x, phd in robot learning.
Katılım Ekim 2018
470 Takip Edilen168 Takipçiler
@roblee_rl By the way just want to what arm is this because I see generalist also uses this arm.?
English
@roblee_rl Is it pose estimation of bottle and picking it up or grasp pose prediction
English
Rob Lee retweetledi
Seeing a model work like this for the first time is such a good feeling
Rob Lee@roblee_rl
a good model trained on even a simple task with a tiny amount of data feels mesmerising, no matter how many times you see it.
English
@sentientcar We use both! This arm is 7dof, has a larger workspace, and a higher payload, which is nice for certain applications
English
@roblee_rl Nice is there a reason you prefer this arm compared to something cheaper like the yam arms ?
English
@carlosdponx @randallmbriggs @GoingBallistic5 Westwood Robotics BEAR actuators also have liquid cooling, used in their humanoids
English
I’ve been saying it’d be cool to see liquid cooling for actuators on humanoids, looks like that’s what Honor did?
@randallmbriggs @GoingBallistic5
RoboHub🤖@XRoboHub
Ran 21 km (13.1 miles) — and the motor was still cold. That’s the detail that matters. 🤖 Honor was the clear dark horse in this year’s robot half marathon. They swept 1st, 2nd, and 3rd, and also posted a strong top-6 finish overall. What stands out to me is that this was not just about bigger motors, or a gait tuned for long-distance running. They seem to have solved something more important — cooling. In a post-race interview, Honor engineers said the robot used liquid-cooling tech adapted from Honor smartphones, with cooling lines running deep into the motor system to carry heat away. Some reports added more detail: the setup used two high-speed micro pumps, with flow rates reaching up to 6 liters per minute, giving the system enough cooling capacity to handle sustained lower-joint motor load. That matters because once a robot starts overheating, output drops, stability goes with it, and the whole run can fall apart fast. And that’s exactly why this detail is interesting. Of course, that does not mean Honor has already surpassed teams like TienKung or Unitree across humanoid robotics as a whole. What it does suggest is that for the marathon task, they built a very strong system solution. And honestly, that alone is already a useful case for the industry. The bigger trend is moving fast. Last year, TienKung won in around 2 hours 40 minutes. This year, the winning time dropped to 50 minutes 26 seconds. Last year, most robots were still fully remote-controlled or only semi-autonomous. This year, around 40% were running with a much higher level of autonomy. So to me, the real signal is not just that robots got faster. It’s that the field is now moving past raw speed, and into the harder problems: autonomy, stability, and system reliability under load. If the pace of progress stays anywhere close to this, then next year’s race should be even more worth watching.
English
@ed0henderson Or more simply, the loss might look like it has plateaued, but the model might still be tweaking the smaller precise parts of the movements.
English
@ed0henderson Imitation learning is weird because there's no great way to pick checkpoints other than eval perf. It's hard to pinpoint overfitting since human demos are noisy/multimodal/not iid. Especially with small datasets the val set might not be perfectly representative either.
English
Look ma no hands! 👋
Starting very small:
- Policy: ACT (Action Chunking Transformer, chunk_size=100)
- Dataset: 20 teleoperated demos (~18k frames) of a human using two SO-101 arms to pick up two Jenga pieces and stack them into a cross.
- Loss plateaued by ~10k steps. 22k training steps on a @modal H100 80GB
#robotics #embodiedAI #physicalAI #robot @ben_giudice
English
Our 3D Vision team (3DGR) is releasing Raiden — a data collection toolkit for YAM robots.
Built for scalable, high-quality data: supports leader–follower + SpaceMouse teleop, multi-camera setups, and modern stereo depth (incl. TRI learned stereo).
tri-ml.github.io/raiden/
English
@Goodeat258 Nice, thanks! Do you plan to release the code for the robotics experiments? Curious how you create the positive/negative samples for imitation learning (since theres only one label per conditioning)
English
We’ve released the code for Drifting Models :)
Includes full training, inference, and pretrained weights.
Curious to see what people build on top of this.
github.com/lambertae/drif…
English
@asimovinc This might be useful, you only need a few simple terms for nice gaits (single foot contact, airtime, simple penalties) arxiv.org/abs/2404.19173…
English
@JieWang_ZJUI Interesting. I guess their contribution is more around training data/recipe?
English
Right, that section shows averaging the outputs of a flow policy doesn't hinder performance much. They also show a figure with very minor spread of modes. In my experiments I found similar behavior, but there are definitely states where diffusion will output multiple modes. In most cases though, you can still get good success rate while collapsing modes, because you will often move to a state with less action ambiguity. (imle-policy.github.io) It's highly dependent on task and dataset though
English
@roblee_rl I think Section 3.2 of the paper is saying that, given the current scale of training demonstrations, generative policies don't actually learn multimodality. They can in principle, but they don't because the amount of data is very limited, right?
English
So the only way to learn multimodal behaviors is to tokenize actions and then use cross-entropy loss?
Chaoyi Pan@ChaoyiPan
Generative models (diffusion/flow) are taking over robotics 🤖. But do we really need to model the full action distribution to control a robot? We suspected the success of Generative Control Policies (GCPs) might be "Much Ado About Noising." We rigorously tested the myths. 🧵👇
English
Great work! Really interesting paper. I'm curious about what you think about recent non-iterative generative policies (C1+C2) like arxiv.org/abs/2502.12371 and arxiv.org/abs/2510.12483. These methods are basically regression but with additional mechanisms that encourage better use of the noise space. It seems that either C1+C2 and C2+C3 can work well, but I wonder about the trade offs.
English
Generative models (diffusion/flow) are taking over robotics 🤖. But do we really need to model the full action distribution to control a robot?
We suspected the success of Generative Control Policies (GCPs) might be "Much Ado About Noising."
We rigorously tested the myths. 🧵👇
English
This might be the best humanoid jog I've seen yet. Doubt this is pure RL, I'd love to know the curriculum here.
Brett Adcock@adcock_brett
English
@tonyzzhao @chichengcc Amazing work! Very curious about the training objective 👀
English
We recycled the ACT name from ALOHA since we love the "Action" part. But the architecture and training objective for ACT-1 is neither the original ACT, nor Diffusion Policy from @chichengcc!
Qianzhong Chen@QianzhongChen
So excited to see that simple but elegant ACT structure has potential to scale up! Congrats @sundayrobotics team!
English
昨日、博士論文の審査を無事に通過したので、CMUでの博士課程の5年間を振り返る記事を書きました! sff8.hatenablog.com/entry/2025/10/…
日本語