Vikash Kumar

Building Human-embodied Intelligence. CEO @MyoLabAI | Sr. research scientist @OpenAI @GoogleAI @AIatMeta | @berkeley_ai @UWcse #MuJoCo | Ad. Prof. @CMU_Robotic

[Vikash Kumar @Vikashplus]

📢Life is a sequence of bets – and I’ve picked my next: @MyolabAI
It’s incredibly ambitious, comes with high risk, & carries unbounded potential. But it’s a version of the #future I deeply believe in.

I believe:
➡️AI will align strongly with humanity - coz it maximizes its own growth & impact
➡️It will transform the world as profoundly as the internet
➡️Like the internet, it will ultimately disappear into the background of our daily lives

Most of what we see today are transient wins - short-term products riding the first waves of capability. Not transients, I’m betting on the signals that will endure.

Just as the cellphone became the personal gateway to the internet era, I believe the future of AI will be 𝐩𝐞𝐫𝐬𝐨𝐧𝐚𝐥𝐢𝐳𝐞𝐝, 𝐜𝐞𝐧𝐭𝐫𝐚𝐥𝐢𝐳𝐞𝐝, & 𝐝𝐞𝐞𝐩𝐥𝐲 𝐡𝐮𝐦𝐚𝐧-𝐜𝐞𝐧𝐭𝐫𝐢𝐜. The interface—the #canvas—of this era is still waiting to be defined.

With MyoLab, I’m placing my bet on the 𝐥𝐢𝐟𝐞𝐥𝐢𝐤𝐞 𝐡𝐮𝐦𝐚𝐧 𝐝𝐢𝐠𝐢𝐭𝐚𝐥 𝐭𝐰𝐢𝐧 as that interface.

We’ve assembled a world-class team with the conviction and grit to make this future real. We’re building a new kind of AI: embodied, personal, and lifelike. Most already believe lifelike digital twins are inevitable. We’re just accelerating the timeline.

Today, we’re releasing an early research preview of the first instantiation of #HumanEmbodiedIntelligence at https://t.co/LxNB3aCEhC

We’d love for you to try it and share your feedback.
𝐓𝐡𝐢𝐬 𝐢𝐬 𝐦𝐲 𝐛𝐞𝐭. 𝐖𝐡𝐚𝐭’𝐬 𝐲𝐨𝐮𝐫𝐬?

[myolab.ai @MyolabAI]

All forms of intelligence co-emerged with a body, except AI

We're building a #future where AI evolves as your lifelike digital twin to assist your needs across health, sports, daily life, creativity, & beyond...

https://t.co/QL3o9YxZYz ➡️ Preview your first #HumanEmbodiedAI

[Vikash Kumar @Vikashplus]

@soaresexis @JieWang_ZJUI @danfei_xu Reality

[Vikash Kumar @Vikashplus]

@MyoSuite is rapidly being the frontier of #ML community developing the next era of Reinforcement Learning.

MyoSuite captures humans morphology in functional tasks. Beyond progress in algorithms, solutions to the tasks presents the potential of being significant impact on life.

[Jusuk Lee @jusukle]

Flow policies are a powerful policy class for continuous-control RL: they represent expressive, multi-modal action distributions, and they train by simple supervised regression — sample improved target actions, then distill.

But in online MaxEnt-RL, one question decides everything:

🎯 Where should the supervision come from?

The usual answer is global importance sampling: sample from the policy, reweight by Q-values, distill. It works only when the proposal can reach high-value regions. In high-dimensional action spaces, that overlap disappears — the proposal misses target-relevant actions, importance weights collapse, and supervision goes sparse.

We introduce FLAG: Flow Policy MaxEnt-RL by Latent-Augmented Guidance ⬇️

🔷 Localize improvement — condition both the proposal and the target on the same flow latent z, so importance sampling happens in a shared local region with real overlap, exactly where improvement occurs.
🔷 No BPTT — update the flow by distilling onto improved action labels, never by differentiating through the flow ODE.
🔷 Principled — a latent-augmented z-MDP with proven Q-function consistency: optimizing the local region is the same problem as the original MDP.
🔷 Provable — a conditional monotonic-improvement guarantee, SAC-style.
🔷 Scales — MuJoCo → DMC Dog → MyoSuite at low GPU cost, robust even at N = 2 importance samples.

🌐 Website: https://t.co/HIbvXQNYL0
📄 arXiv: https://t.co/dRkVzzoBuB
💻 Code: https://t.co/l4JQFlhdj5

[Vikash Kumar @Vikashplus]

Robotics is community built on critical fundamentals like these - analytical IK for arms.

[Siddhartha Srinivasa @siddhss5]

Fifteen years ago, I had the privilege of working with Rosen Diankov at CMU on his PhD thesis. The capstone was IKFast — for a generation of roboticists, the definition of what analytical inverse kinematics could be.

Today, I'm excited to release the next chapter: ssik.

1/

[Vikash Kumar @Vikashplus]

@RoboPapers @micoolcho @chris_j_paxton @DJiafei @BitRobotNetwork @SharpaRobotics @LightwheelAI @hq_fang @sanatem @Noriaki_Hirose @gao_young Back in 2018, @shaneguML and a few others proposed #Origami as an internal moonshot challenge for robotics. It has it all :
- dexterity
- long horizon planning
- flexible object manipulation
- infinite task diversity

[Vikash Kumar @Vikashplus]

Shout out to a new pre-trained vision model for robotics that comes close to and outperforms prev works from our group - R3M (w @SurajNair_1), VIP(w @JasonMa2020), VC1 (w @aravindr93), etc.

[Jusuk Lee @jusukle]

Are you still running your robot policies on vision encoders trained purely on static images?

Nowadays, the standard practice in robot learning is to plug in powerful vision models like CLIP, SigLIP, or DINOv2. This inherits a quiet, convenient assumption: “Let mainstream computer vision handle perception, and the downstream policy will figure out the dynamics.”

But let’s be real for a moment. Is this truly the best we can do?

We introduce DynaFLIP: Rethinking Robotics Perception via Tri-Modal-Dynamics Guided Representation.⬇️

🔷 Dynamics upstream: we push motion understanding into perception.
🔷 Tri-modal-dynamics supervision: image transitions × language × 3D flow, fused via simplex-volume alignment (260K trajectories from robot & human video)
🔷 Transfers everywhere: a visual backbone for diverse policies (MLP, Diffusion Policy, VLA)
🔷 +22.5% over the strongest baseline (DINOv2, SigLIP) under real-world OOD
🔷 Open-Source & easy to use

🌐 Website: https://t.co/I3uKpAZ975
📄 Paper: https://t.co/jHAweJBreK
💻 Code: https://t.co/yUueJ1xxJL
🤗 Hugging Face: https://t.co/jqLzJFHvMI

[Vikash Kumar @Vikashplus]

Main criticism with simulations are their difficulty+gap in capturing real world diversity. Progress in Generative simulations speeding up fast 💪

The focus needs to move from photorealism to physics & forces -- the language of physical world.

[Genesis AI @gs_ai_]

We are back again :) After three weeks of quiet building.

Introducing Genesis World 1.0, our latest simulation platform, the second release in our full-stack suite. Open-sourced.

Robotics is still bottlenecked by the 1× speed of the physical world. Every model, checkpoint, and data recipe eventually needs to be tested on physical hardware, slowly, expensively, and with limited coverage.

One hour in reality can become 100 days in simulation. That is how robotics model iteration moves from a wall-clock bottleneck to a compute problem.

To make this work, simulation has to be both fast and trustworthy.

Over the past year, we rebuilt the entire stack: a GPU-accelerated cross-platform compiler, penetration-free multi-physics contact solvers, unified rigid and deformable physics, and a photo-realistic renderer purpose-built for physical AI applications.

We built Nyx, a high-performance path-traced rendering engine for robotics application.

Genesis World 1.0 achieves near realtime performance with our latest development for penetration-free IPC solver, supporting various types of deformables beyond rigid bodies. It supports contact-rich, dexterous manipulation simulation across different embodiments: unitree, sharpa, wuji, genesis hand and various types of grippers.

Under the hood is Quadrants, our effort in pushing forward cross-platform GPU-accelerated computation. Quadrants started as a fork of Taichi, and we rebuilt most of the critical parts for optimizing simulation workloads, giving 10x faster launch time and up to 4.6x runtime performance compared to the initial Genesis release.

Together, they bring us to an unprecedentedly low sim-to-real gap, enabling zero-shot real-to-sim model evaluation and much faster iteration of GENE.

All available today.
Genesis World 1.0: https://t.co/aknCM3eqws
Quadrants: https://t.co/uXqPNI4cb6
Nyx: https://t.co/R8j0djqGnV

[Raj Patel @ CVPR @babugi28]

Today, Human Archive is announcing our $8.2M seed round to model human embodied intelligence.

Despite decades of research, we still barely understand ourselves. Our goal is to learn how humans interact with the world, and over the past 6 months, our team’s made enormous progress toward that alongside leading AI labs.

learn more @TechCrunch
https://t.co/faLhyVBjl1

[Vikash Kumar @Vikashplus]

Race for humanoids continues where 4 bars and 2 wheels can be equally effective!!

[Space and Technology @spaceandtech_]

This is evoBOT, a robot helper developed by Germany’s Fraunhofer Institute for Material Flow and Logistics.

It can grasp and carry goods to support cargo workers in transporting packages.

evoBOT can also move smoothly across uneven terrain, including bumpy surfaces and sloping ground.

[Vikash Kumar @Vikashplus]

A great reminder that -
robot learning = data+algo+scale

[Stepan Feduniak @FeduniakS]

Spent last week benchmarking policy speedup methods. Then we just collected faster data and it beat all baselines...
Although obvious, but turns out first step to speed up your policy is … collect faster data.

[Vikash Kumar @Vikashplus]

✈️In ☀️California (until Friday)

Hit me up if you would like to meet. DMs open.

[Vikash Kumar @Vikashplus]

@abhishekunique7 The observation distribution is quite different between sim and real. I’m curious why it’s reasonable to freeze the encoder, reward, and value function on the prior distribution?
Aren’t we introducing inconsistencies?

[Vikash Kumar @Vikashplus]

@wenlong_huang And keeping the dynamics and policy distribution well aligned with each other.

(Good dynamics model where policy is exploring states unhinged is disaster)

[Vikash Kumar @Vikashplus]

@abhishekunique7 Do we need to know the task distribution, or can it be trained agnostic to the distribution ahead of times?

[Vikash Kumar @Vikashplus]

@macdonaldncode Imagine getting an intern and not being able to talk to him/her, it will become very hard to train them.

Humans have always aligned with the changing workflows. This is our superpower. And this time will be no different.

[Vikash Kumar @Vikashplus]

text2motion often struggles from physically inconsistent motion. However, the whole body controllers for G1 has gotten so robust that it’s finally becoming possible to connect the two.

Next frontier - task driven contextual motion generation & real time execution - can provide a great interface for training robots on the job.

[Unitree @UnitreeRobotics]

Voice‑driven, real‑time arbitrary action generation😁
Using external voice commands, G1 is directly controlled to generate a wide range of actions in real time.
This video was recorded in a single take, with on‑site audio recording.
Because the actions are autonomously generated by AI in real time, there may be slight latency, and the smoothness of the movements may be somewhat reduced.

[Vikash Kumar @Vikashplus]

SCALING ISN’T EVERYTHING

Another tiny model breaking the rule.
-trained on less than 1/1000th of the data
- can be trained in a single day with <1000 USD

Human knowledge base ca be compressed & retrieved much tighter than LLMs do today.

[Vikash Kumar @Vikashplus]

@lukas_m_ziegler @OliveRobotics The latency as observed from the visualization seems a bit high, no?

[Vikash Kumar @Vikashplus]

@ErenChenAI It’s very disturbing that there isn’t a kill switch on the robot that can be quickly pressed !

[Vikash Kumar @Vikashplus]

Whole body controllers - effective with contact rich behaviors - are the unsung HEROs of robotics🦸‍♂️

Without them all we will have - is a bunch of over powerful pincers picking & placing tiny objects on the table.

(a bit harsh but true)

[Alberto Rodriguez @_albertorod_]

You can’t lift a fridge with just your hands. Your whole body needs to conform to its shape, and bear the load between your arms and torso.

Here, @BostonDynamics' Atlas uses proprioception to manage the whole-body interaction and adapt to a shifting 100+ lb load. Enabling this type of high performance manipulation is exactly why we walked away from what was arguably the world’s best implementation of MPC for humanoids, and shifted entirely to RL without looking back.

This level of whole-body controls is a fundamental building block of physical intelligence and key to the value proposition of humanoids.

More technical details in:
Blog: https://t.co/oIRjVfh7jJ
Behind the scenes video: https://t.co/LgaImMAyhX