Genesis-Embodied-AI open-sources Genesis World 1.0, a GPU-accelerated physics simulator running up to 100x faster than real time

VIEWS37KLIKES120

Cool gets cooler

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://github.com/Genesis-Embodied-AI/genesis-world Quadrants: https://github.com/Genesis-Embodied-AI/quadrants Nyx: https://github.com/Genesis-Embodied-AI/genesis-nyx

26d37K12034

BOOKMARKS38

Kun Lei@kunlei15

After reading Genesis World 1.0 and GENE-26.5, one point I strongly agree with is that a simulator/world model is critical for policy evaluation, especially for general robot foundation models.

Robotics is different from static prediction. Once a policy acts, it changes the world. The next observation depends on contact, dynamics, friction, object motion, latency, embodiment, and many tiny physical details.

So open-loop dataset metrics are not enough. A robot policy needs closed-loop evaluation: roll out the policy, observe the consequences, and ask whether the world moved toward the goal. But a good world evaluator can do more than evaluate. It can guide policy improvement.

Offline RL is a good analogy. In offline RL, we want to improve a policy from a fixed dataset, or a slowly updated data flywheel, without frequent online access to the real environment. The data may come from a behavior policy, human demonstrators, or teleoperators. The goal is not only to imitate them. The goal is to improve beyond the behavior represented in the dataset.

The hard question is: once we update the policy, how do we know the new policy is actually better? The dataset only tells us what happened under the old distribution. It cannot directly tell us what will happen when a new policy takes new actions, enters new states, or composes skills differently.

A reliable forward dynamics model can close this loop. It can simulate consequences, evaluate counterfactual action sequences, provide multi-step feedback, and guide policy updates before real-world execution.

This is also the recipe we use in RL-100 / Uni-o4: multi-step policy iteration guided by a forward dynamics model. The model is not only used to predict the future. It is used as an evaluator and improvement signal for the policy.

For robot foundation models, I think this policy/world-model loop will be central. As the robot data flywheel starts turning, the model sees more information about the physical world: object properties, embodiment motions, contacts, task structure, temporal dynamics, and the relationship between language, observation, and action.

One part of the system learns the mapping from observation and language to motion. We call this the policy.

Another part learns how the physical world evolves under actions: how objects move, how contacts happen, how errors accumulate, and how long-horizon consequences unfold. We call this the world model.

These components can be integrated or separated architecturally, but they should reinforce each other.

A better policy produces better interaction data. A better world model provides better evaluation. Better evaluation enables better policy updates. Better policies reveal more of the physical world. This loop may be one of the key ingredients for scaling robot foundation models. Behavior cloning gives us a strong starting point. Teleoperation gives us high-quality demonstrations. Large datasets give us broad coverage. But to go beyond imitation, robots need a way to evaluate and improve policies after the initial dataset.

That is why a world model is so important. A good world model is not just a simulator of pixels. It is not just a generator of synthetic trajectories. It is a scalable evaluator of counterfactual futures.

And once the evaluator becomes good enough, policy learning can move from static imitation to iterative improvement. That may be the bridge from robot models that imitate human data to robot models that improve beyond it.

Genesis AI@gs_ai_