dave

robotics enthusiast, @GeorgiaTech alum

[dave @davidhe137]

such an inspiring blog post. there is still so much you can learn from sim!

[Seohong Park @seohong_park]

Behavioral cloning mystery

https://t.co/VqxzvcfGSx

I wrote a new blog post about "mysteries" in behavioral cloning that appear with real-world robot data (e.g., overfitting is "good"). I also tried to demystify them and shared my thoughts!

[Woo Chul Shin @woochulshin1726]

Train on static objects. Grasp moving objects. 🤖

Many visuomotor policies assume static objects. Real deployments break this: parts drift on conveyors, fruits sway in the wind. Meet SIDO (Static In, Dynamic Out), taking a policy trained only on static demos to moving objects.

[dave @davidhe137]

@DozenDucc a very possibly shortsighted question: What are the benefits of LLMs/code-as-policy for robotics over imitation learning besides high-level planning/zero-shot capabilities?

[Zhenyang Chen @DanielZhenyang]

Human data is the cheapest source of whole-body mobile manipulation demos. Turning it into robot data is the hard part.🤖

WARP retargets OFFLINE human motion into replayable robot actions in closed form — precise, consistent, faithful to whole-body intent.
🌍 https://t.co/gB77gIQWKp

[Yuke Zhu @yukez]

Today, we're introducing SimFoundry, our real2sim2real framework at NVIDIA GEAR that automatically turns real-world scenes into simulation-ready worlds from a single image or video.

Website: https://t.co/JB3kf3GlYm
Paper: https://t.co/pVhE1qWXtU

This work marks a major step for our team toward leveraging simulations and synthetic data for foundation model training and systematic policy evaluation at scale. Code will be open-sourced soon. Stay tuned!

[Ajay Mandlekar @AjayMandlekar]

We have released the SimFoundry code! Excited to see how others will use it.

Check it out here: https://t.co/qpcUncxBm6

[dave @davidhe137]

@Ryan_Resolution @IsaacSin12 @makermodsai very cool! how does this work?

i’ve been having some trouble getting consistent policy performance when deploying to many different arms (same make, maybe just uncalibrated). any advice for that?

[dave @davidhe137]

@yuleiqian sorting runs at 50hz and spinning disk runs at 40hz

[dave @davidhe137]

train-time rtc is a beast

[dave @davidhe137]

@gundala_geledek @chen_bingh85992 we have a leader-follower setup with almost no lag, credit to @rbansal22. iirc, we use MIT mode on piper_ros.

Also should be mindful of teleop quality (have some failures/recovery but not as sloppy as shown here)

[dave @davidhe137]

remote inference should be under 100ms for decent GPUs (l40s or better), and networking should not be too bad either if the GPU is in the same region as your robot. For reference, these policies are running at 40-50hz with action horizons of 20.

You can try Armory, our project which supports low-latency remote inference on @modal: https://t.co/5ssN33IxCq

Will try to merge support for train-time RTC in the next few days. Happy to answer any more questions in thread or DMs!

[dave @davidhe137]

@yuleiqian we finetuned pi05 on our own data, more details here. No custom planning or control for RTC, its just standard action chunking on the AgileX Piper joint states.

[dave @davidhe137]

We collected ~60 episodes per task (~10 pick places per episode) but mainly because we are deploying on many different robots with slight differences (so we collect on many robots). You can probably collect ~30 episodes and see good results.

Data is https://t.co/L0Ku0toSSb

Code is https://t.co/9UG2MKwC5d

[dave @davidhe137]

@cyrux004 haven’t tried on so-101 but i’ve seen other posts do it: https://t.co/JV4JCXBDUB

we use pi05 and you can refer to the original paper https://t.co/UKfElsdMGp

[Jack Vial @jackvial89]

60hz! real time chunking on an so101 with LeRobot. Not looking too bad. Bit of jitter throughout but no mode switching across chunks

[dave @davidhe137]

@kjsRobotics RTC is just inpainting the prefix of the generated chunk with the actions that will be executed during inference. You can do guided flow-matching at inference time or train your model to inpaint. See https://t.co/UKfElsdMGp

[dave @davidhe137]

@chen_bingh85992 also, inference code coming soon to Armory! (need to clean up)

[dave @davidhe137]

We collected ~60 episodes per task (~10 pick places per episode) but mainly because we are deploying on many different robots with slight differences (so we collect on many robots). You can probably collect ~30 episodes and see good results.

Data is https://t.co/L0Ku0toSSb

Code is https://t.co/9UG2MKwC5d

[dave @davidhe137]

Haven’t compared to test-time RTC, it’s a bit hard to both implement and debug🥲

The policy here is actually trained on 30hz but i just bumped up the execution hz to 50 without issue. Since this is just 1 robot, inference just runs as fast as it can (on Armory!) and the prefix is just inference delay. Was pleasantly surprised that it just works out-of-the-box!

[Ryan Punamiya @ryan_punamiya]

📢 Call for Papers: Scaling Human2Robot @ CoRL 2026

Human data is slowly becoming a central part of scalable robot learning. But what can we expect as it scales?

Join Scaling H2R @ #CoRL2026 🤖
⏰ Submit by Oct 7 (AoE)
🌐 https://t.co/tetnaouiY4

[dave @davidhe137]

@LAKSHYA49332151 @rbansal22 We use wifi in our experiments. The policy takes a wrist and a top view camera, which are both 224 x 224 x 3.

[dave @davidhe137]

Check out our new work on serving robotics policies from the cloud!

[Rohan Bansal @rbansal22]

🤖 Scaling deployment in robotics can be difficult and expensive. How can we properly serve a VLA/WAM to 10+ robots concurrently from a single GPU?

Introducing Armory, a system for serving large robot policies to a fleet of robots from the cloud/edge. Through real-time scheduling of action chunks, Armory boosts task throughput over baselines by up to 18% on heterogeneous fleets in the real world.

🧵1/6