vic_ssbb

For those interested in robot locomotion (humanoids, quadrupeds, but also robot control and learning in general): I created a discord server so we can exchange: discord.gg/Nn6PAyAg

@GoingBallistic5 I am curious to know why they got rid of all the parallel linkages (ankles, knees and torso ? Maintenance ? Simulation ? Both ?

If this is someone's idea of a cruel joke, it isn't funny

@DFintelligence À Telecom Physique Strasbourg, il y a maintenant un cours de Reinforcement Learning et simulation dans la filière robotique. Cela s'ajoute aux cours de DL, Signal Processing, Vision 3D déjà présents

Il y a des écoles d’ingénieurs en France qui ont lancé des filières de AI Robotics ? Ou pas encore ?

@GoingBallistic5 It was already the case for H1, at least H1-v2

The Unitree H2 Knee is now remotely actuated The knee actuator is higher, making its mass more proximal, reducing hip torque It is just below the hip R actuator and protrudes slightly inboard. This allows the tie rod to the knee joint to remain centered along the femur

@KyleMorgenstein @BodyMindAI I see no reason why the pin joint (actually two revolute slightly displaced) would not work in USD. Maybe its the joint limits that arent converted to USD. Btw, I wrote a paper explaining what I think is the good way of controlling this mechanism: arxiv.org/abs/2503.22459

@vic_ssbb @BodyMindAI There are MJCF/URDFs for a pin joint form, but when IsaacSim converts them to USD I don’t think it does a great job, and the company hasn’t put out their own USD publicly. The pin joint model is remapped through the low level jacobian which is where the singularity issues pop up

Yep! It's happening. And DM me if you would like a quote. The price is.... Surprising.

@KyleMorgenstein @BodyMindAI Does the Isaac Sim USD models the closed linkages ? Or it is a fully serial model ? It is interesting. As far I as I know, the way the low level works for these 2Dof ankles is not never explicited for these commercial robots.

@vic_ssbb @BodyMindAI which is a shame because it’s otherwise a great robot, we have a few at UT that I really wish I got more use out of!

@KyleMorgenstein @BodyMindAI What was the issue with the ankles ?

@BodyMindAI How are the ankles on K1? This was a challenge for the T1s we have. Would be very excited to hear a price point for a model with low level control.

@KyleMorgenstein What rewards are you usually using for locomotion, instead of the Isaac lab examples rewards ?

to converge quickly (<1500 iters) but the rewards are finicky and don’t transfer well. by comparison, the rewards I use usually require ~30k iters but transfer to hardware beautifully and are easy to tune for different behaviors. at some point you just need more gradient updates.

read the holy scripture, bröther from arxiv.org/abs/2109.11978, increasing parallelism and batch size results in better policy gradient estimates so more efficient training. you can find the Pareto frontier for performance given the complexity of the task and compute budget.

@ChongZitaZhang But contact mesh can be simplified using primitives ? Not sure this is the reason. In my experiences contact meshes changes a lot the policy in a way that is difficult to understand (at least in PhysX simulators)

@vic_ssbb too big to test. torque limits on certain dofs are also too small. the contact mesh is also more complicated.

A student just trained this within a day, no tedious tuning, no sim2real tricks, not even with sys-id. Worked in the first trial on the real robot. This explains the many recent impressive demos on G1 robot -- just the hardware. Still sim2real gaps on ankle and waist dofs tho.

@KyleMorgenstein Would definetly be interested.

I wish there was a good venue to write/present about RL “tricks”. how PD gains affect action scale, how to tune reward functions, actor STD, etc. there’s good intuition for all of it, grounded both in learning theory and robot dynamics, but I don’t often see good explanations.

@KyleMorgenstein @Mankaran32 Can you share some additional information/ressources related to this armature issue please ? Recently had issues with armatures on my Isaac models.

@Mankaran32 just a legacy thing, for well configured robots, IsaacSim sim2real is a dream. RAI/ETH/BD/UT all use IsaacSim for sim2real. but it currently has an issue with armature and has some momentum conservation issues earlier this year. they’re minor and only pop up in edge cases tho.

Nvidia will also have monopoly in simulators. Isaac is undefeated and very difficult to dethrone. If humanoids are a ~$3T opportunity then this is what will power all of them.

@DFintelligence @PeugeotFR @renault_fr Je suis ingénieur dans le domaine et je suis intéressé pour participer à un projet de ce genre. Si quelqu'un est intéressé, envoyez-moi un DM.

Je préshot : la France doit avoir un fleuron pour créer des humanoïdes avec un prix accessible d'entrée (10k€). Mais pas un projet qui sortirait dans 3-4 ans, il nous faut un prototype en 10 mois. Pourquoi @PeugeotFR ou @renault_fr ne s’y mettent-ils pas ? Ce sont les seuls acteurs que je vois capables de faire un banger. Poke @ClaraChappaz pour pas focus toute l'attention sur les LLMs... Ou alors, booster le programme robotique de @huggingface. Je crois à 100 % aux modèles de robotique open source dont on avait parlé avec @ClementDelangue. (bon sinon donnez-moi un département avec 25 M€ de budget et on le fait si vous voulez je veux juste l'appeler Skynet 🤡.)

@hankyang94 Paper link ?

How robust can model predictive control be if we can solve each trajectory optimization to global optimality? On the contact-rich push-T problem, we show that model-based global optimization is so robust that it never fails, even if the model is not even correct! We achieve global optimality via sparse Moment and SOS relaxations. -- Yes, we managed to solve SDPs online on a robot. Amazing work by @ShuchengK and Guorui Liu.

@IFAC_Control @ChongZitaZhang @brianbdouglas What about iLQR/DDP ?

Let's get ready for the first round of control fun in 2025. 🤩 #cartoon by @brianbdouglas #controlengineering #automaticcontrol

@Mister_Gob T'as pensé quoi des derniers scans chef ?

@I_loves_deep_nn I see, and is the control of the EUV machine the most critical part or ASML just master the whole pipeline better than any other ?

@I_loves_deep_nn Why does TSMC have monopoly on advanced chip when it is ASML that builds the EUV machines ?

@JoeChiiip Quelqu'un aurait cet article complet et pourrais me le partager par MP, svp ?

@albanvln @JeanAbbiateci @lorisguemart Son travail consiste en l'écriture de livres de philosophie des sciences. Son HDR est aussi un de ses livres grand publics. Est ce que ça fait de son poste un "emploi fictif" ? Est ce en dehors des règles ?

@JeanAbbiateci @lorisguemart Étape suivante, son emploi fictif de directeur de recherche au CEA…

Étienne Klein, une thèse constellée de plagiats (par @lorisguemart et moi-même) arretsurimages.net/articles/etien…