Kyle🤖🚀🦭

when you argue with me about control theory this is who you’re arguing with

@LouisLeLay4 @kevin_zakka @ChongZitaZhang No worries I know lol :) I can’t wait to share!

@KyleMorgenstein @kevin_zakka @ChongZitaZhang Sorry man I'm messing with you ahah but super cool can't wait to hear more abt it :)

This is why we need Newton

@LouisLeLay4 @kevin_zakka @ChongZitaZhang when I find the time I’m very excited to play with the depth sim2real so101 envs. I’ve also used that viser setup for many things, it’s great!

@LouisLeLay4 @kevin_zakka @ChongZitaZhang No no I’m no ideologue! I’ve rewritten my whole codebase three times during my PhD for different backends. Right now I just have a lot of load bearing infra + need photorealistic rendering + am soon to graduate so it’s a hard sell atm, but I’ve loved seeing its development!

There’s no good time for a robot to be out of commission, but a week before a deadline trying to shoot the video is certainly the worst. 🥲

@kevin_zakka @ChongZitaZhang you knew what you were doing 😉😘

@KyleMorgenstein @ChongZitaZhang Just because I wanna have fun, they both didn't like the tweet 🧂🧂🧂🧂

Working on the SLAM Handbook has been one of the highlights of my career — I’m grateful to have collaborated with such an incredible group of co-editors and contributors. The handbook is now free and open-source: dspace.mit.edu/handle/1721.1/… [1/n]

@LouisLeLay4 @ChongZitaZhang also you have to cast from warp to torch like shrapnel everywhere

@LouisLeLay4 @ChongZitaZhang :( #migrating-to-isaaclab-3-0" target="_blank" rel="nofollow noopener">isaac-sim.github.io/IsaacLab/devel…

@ChongZitaZhang also related to contact model, causes large non physical spikes, thousands of newtons in a single (sub)step. adding EMA can help.

Find something severe to be improved😅. The link acc fetched directly from physx is noisy and can wrongly terminate environments. I think I will fix that by using finite difference (was actually done in my paper). Also for force measurements I now need to filter and clip it.

As an application of the invariant EKF in GTSAM, we added a state estimator for legged robots, along with some more factor-graph-heavy variants. Try it on your rosbags - especially interested in validating it with non-quadrupeds (bipeds, hexapods, spiders). It takes IMU and feet in body frame as inputs.

Animation 🤝 Robotics ProtoMotions GTC 2026 release — bridging the gap between digital humans and real humanoid robots. Train in simulation. Deploy on hardware. One framework, one codebase. nvlabs.github.io/ProtoMotions

@ChenTessler finally got it on hardware. respect!! super excited to play with it

@KyleStachowicz @KyleVedder @vriishin @saurabhtwq kylebot was my display name for years, never did I imagine “robokyle”; that’s some real innovation right there

@KyleVedder @vriishin @saurabhtwq @KyleMorgenstein robokyle group chat when

would love to follow more people who working in vlm, world models and robotics. any recs ?

@saurabhtwq @KyleVedder and @KyleMorgenstein 2 goats somehow both named kyle

@observie You can also improve energy usage by penalizing foot velocity at initial contact and negative foot force, but then you get into the reward shaping swamp

Energy Optimization in mjlab Quadruped Locomotion During locomotion, motors waste battery energy in two ways: 1. Heat (I²R losses), approx proportional to τ², already penalized in Q1 Hoper's baseline policy via the joint_torques_l2 term. 2. Mechanical power (τ⋅ω), the actual work delivered by the motors. Positive values indicate propulsion; negative values indicate braking/regen. In principle regen could recharge the battery, but Q1 Hoper does not implement this (unclear for other robots). I added an energy penalty of -5e-4 × sum(|τ⋅ω|), penalizing both positive and negative mechanical power, on top of the existing torque L2. Result: The energy-penalized policy discovers a noticeably more efficient gait with zero performance loss. Baseline (only L2 torque term): 5 air / 5 ground frames → 50% air time Energy-penalized: 6 air / 4 ground frames → 60% air time Same episode length, same velocity tracking, improved soft landing, and reduced foot slip.

@ChongZitaZhang releasing the terrains? is this all taka’s terrain generator?

zero shot test passed.

Today marks a major step for industrial automation. ABB Robotics and @nvidia have closed the sim-to-real gap - achieving 99% accuracy with RobotStudio® HyperReality. Design, test and optimize production lines virtually with unprecedented confidence.

Estimation is not hyped up enough but it's very cool

Luckily, hard problems are hard! There will still be much to do, now, and in the future.

I miss working on estimation. If we can figure this RL loco-manipulation thing out, I hope I can find time to work more on estimation. I was supposed to spend my PhD doing estimation, but then I got involved with IsaacGym/IsaacLab and that took over my life.