Cybernetic Physics

Forward deploying robotic intelligence
@speedrun 007

[Cybernetic Physics @cyberneticphysx]

Policies can be seen as actions. In ROS they're called actions, in agents it's skills or tools, but regardless you have an agent that has some certain notion of the skills or tools it can use.

[Cybernetic Physics @cyberneticphysx]

Humanoids are a complicated enough system that require a behavior-level CI/CD. It can live in Git, but behaviors need to be structured in such a way that they are fixed, tagged policies for whatever the underlying things are that you've learned.

[Cybernetic Physics @cyberneticphysx]

3 / Amazon warehouses are so highly automated because Amazon redesigned them until all physical work could be reduced to consecutive pick and place tasks. In datacenters, though, the situation is more complex. Imagine, for example, you have a robot replacing a liquid-cooling line, and you have a hose that is leaking. It's pretty hard to code up a general understanding of that state space in simulation, but the robot needs to absolutely figure out on the spot, "Oh, that's really bad, I can't have coolant leaking onto live electrical equipment."

[Cybernetic Physics @cyberneticphysx]

2 / Another requirement would be for the robots to be "self-healing" in the sense that they have to maintain not only the site but also each other without a human in the loop. Here’s a trained policy in MuJoCo of a Unitree G1 changing the battery of another Unitree G1:

[Cybernetic Physics @cyberneticphysx]

What would it take to run a datacenter "lights out"? @sama's stated ambition is to eventually stand up a gigawatt a week at around $20B per gigawatt yet 90% of datacenter operators say staffing shortages are a critical constraint on building and expanding sites. 🧵

[Cybernetic Physics @cyberneticphysx]

1 / A robot would have to handle rack-and-stack, cabling and cross-connects, hardware swaps, reboots, upgrades, and on-site troubleshooting. Basically all the things currently handled by "smart-hands" technicians in colocation facilities. The datacenter work that has not already been automated requires a human because the physical infrastructure has been designed around one.

Examples:
--> someone has to stand sideways in order to fit in the hot aisle behind the rack and switch cables
--> cage nuts and fiber connectors assume fingers, a screwdriver etc.
--> cables are color-coded

[Cybernetic Physics @cyberneticphysx]

Generating millions of simulated robot episodes should be as easy, scalable, and verifiable as generating LLM rollouts. Only then will we start seeing interesting sim-to-real failures, scaling laws with test-time compute, and eventually some sort of auto-research for physical AI. Slowly and then all at once.

[Cybernetic Physics @cyberneticphysx]

π0.5 being on the level of GPT 3.5 is an absurdly high standard given that the former has been trained on orders of magnitude less data than the latter. For robotics, we will need the equivalent post-GPT 3.5 infrastructure that now makes GPT 5.6 feel inevitable. We need to be able to scale simulation, trajectories, episodes efficiently in the cloud. If you want these models to improve through RL, you need to be able to run huge numbers of robot rollouts, collect the trajectories, evaluate whether the robot actually accomplished the task, and then feed those results back into training. And those rollouts are much more expensive because you're running physics simulation, rendering, perception, the policy, all of these things together.

[Cybernetic Physics @cyberneticphysx]

To go from "you can fine-tune a model to be really good at text-to-SQL" to "you can fine-tune a model to get a perfect score on the Putnam, solve Erdős problems, write SOTA Triton kernels etc.," we needed two things: strong priors (good base models) + mature infra to power scalable, self-improving ML systems. Insert @modal, @togethercompute, @tinkerapi, @anyscalecompute, @PrimeIntellect and many more that abstracted away a lot of the infrastructure pain involved with spinning up compute and sandboxes, running training / inference and generating rollouts at scale. We will see something similar play out in robotics🧵👇

[Cybernetic Physics @cyberneticphysx]

Physics-based world models are not here to replace simulators but they are the missing piece required to scale RL in robotics effectively. Say you have an agentic loop, where a world model's role is to dream up the next state. You still need the simulator in the loop to anchor the planning. The simulator is the more expensive, more verifiable step, whereas the "dreaming" loop becomes the quicker (yet necessary) approximation.

[Cybernetic Physics @cyberneticphysx]

The physics simulator is a verifiable substrate, at least to the level that the simulation and the physics engine reflects reality. The fact that there is non-determinism (you can't really project that far into the future even in a physics simulator) does not mean that you cannot project multiple possible futures with some variation.

[Cybernetic Physics @cyberneticphysx]

Industrial robots have existed and been operating dependably for decades but we're currently seeing this inflection point where world models are becoming impressively good at allowing a robot to embody more intelligence than just a "dumb" industrial robot of the past. From the human manager perspective, the robot still has to follow the same rules in terms of safety, predictability, and explainability. But it can also have a level of understanding and creativity that is necessary for the robot to become a useful member of the factory floor, the datacenter, the shipyard.

This is what we do at @cyberneticphysx.

[Cybernetic Physics @cyberneticphysx]

The ChatGPT moment for robotics will require an autonomy architecture that is well beyond scripted autonomy. In classical autonomy, we have things like state machines and behavior trees. But with models like @nvidia Cosmos 3, we can have an omni-model dream up the next state and then use that to execute an action.

At @cyberneticphysx we are combining the two approaches for our deployment and autonomy stack. In an industrial setting, we use the world model to generate executable specifications (a behavior tree), where the kind of atomic actions in that structure or constrained grammar are primitives that the robot knows how to execute because they're things that we've already trained in simulation for. So either we have VLAs for specific manipulation activities, or a locomotion policy, or even, for example, for a task like welding, it can be a classical control algorithm. Then the robot has to piece all of those things together and then adjust on the fly if something goes unexpected.