The Physical AI Toolbox: Six Free Simulators, a $249 Hardware Ladder, and What the 2026 Papers Admit

On 1 June 2026, a robot walked onto the GTC Taipei stage: a Unitree H2 Plus body, five-fingered tactile hands from Sharpa Wave, an NVIDIA Jetson Thor where the head should be. About six feet, 150 pounds, 75 degrees of freedom.

NVIDIA said it was not for sale. They called it the GR00T reference humanoid — a standard machine for universities to do research on, with the Allen Institute for AI, ETH Zurich, the Stanford Robotics Center and UC San Diego named as first partners.

The interesting part was not the body. It was that nearly everything running inside it — the simulator, the training framework, even the foundation-model brain — is something you can download onto a laptop today. The price is zero, and the licences allow commercial use.

Six tools, and what each difference actually changes

First, a trap to avoid. Search “robot simulation software” and most results are offline-programming tools for industrial arms — software for previewing the path of a robot that is already programmed. Physical AI simulators are a different species: they run thousands of virtual worlds in parallel so a robot can teach itself.

Tool	Maker / cost	What it is for	What the difference changes
Isaac Sim / Isaac Lab	NVIDIA / open source	GPU-parallel physics + photoreal RTX rendering; large-scale RL	Practically the only choice when the policy learns from camera pixels. Without an RTX GPU it will not even start
MuJoCo	Google DeepMind / Apache 2.0	Precise contact physics; MJX runs on GPU/TPU	When fine contact decides the task — a hand closing on a cup — this is the accurate one. Renders plainly, so weak for vision learning
Genesis	CMU-led research team / open source	Pure-Python ultra-fast physics; 4D scene generation from text	“Months of training in hours” territory — but the 430,000× real-time figure is one specific benchmark scene, not your workload
Gazebo (Harmonic)	Open Robotics / open source	ROS 2-native integration, sensor plugins, large outdoor worlds	For validating a whole ROS stack — delivery robots, autonomy. Slower for learning than the three above
Open X-Embodiment	21 institutions / public dataset	1M+ real robot trajectories, 22 robot types	Removes the cold-start data problem. The finding: other robots’ experience helps even when your robot is not in the list
Isaac GR00T N1.7	NVIDIA / Apache 2.0	Open VLA foundation model for humanoids	You no longer build the brain from scratch — and commercial use is allowed, so it can ship inside a product

The question every beginner asks: which one should I pick? The honest answer is that practitioners do not pick one.

One piece of history worth knowing: MuJoCo began in 2012 as Emo Todorov’s project at the University of Washington, was acquired by Google DeepMind, and went fully Apache 2.0 in 2022. Software that once required a paid licence key is now one pip install.

If the frameworks are free, why did NVIDIA need 105 authors?

Isaac Lab’s official paper (arXiv:2511.04831, November 2025) lists 105 NVIDIA-affiliated authors — it reads like a white paper. What it emphasizes is not speed but integration: actuator models, sensors simulated at different rates, data-collection pipelines and domain-randomization tools in one platform. Translation: until now, researchers burned their time gluing those pieces together themselves. The paper also announces Newton, a differentiable GPU physics engine — when that lands, you will be able to differentiate through the simulation and optimize directly.

Hardware — where the money starts

Free software means the differentiation moved to hardware and data. The hardware sorts into three layers.

1. The chip that rides inside the robot

Product	What you get	Price (official)	What the difference changes
Jetson AGX Thor dev kit	2,070 FP4 TFLOPS, 2,560-core Blackwell GPU, 128 GB LPDDR5x, 40–130 W	$3,499 ($3,429 at 10+)	First generation that can run a VLA model inside the body. The 128 GB is the point — the whole model fits on board, so no cloud round-trip
Jetson AGX Orin	1/7.5 the AI performance of Thor	varies by module	The class Boston Dynamics’ Spot actually uses. Plenty for locomotion policies; cramped for large language-vision models
Jetson Orin Nano Super	67 TOPS (1.7× its predecessor)	$249	The realistic starting point for individuals and classrooms — enough to run small VLMs and vision transformers on-device

2. The robot itself — entry now starts at $100

The friendliest change is the entry price. The SO-101 arm from Hugging Face and RobotStudio is 3D-printable from public files: roughly $130 if you source the parts yourself, or $220–240 for the SO-ARM101 dual-arm kit. It plugs straight into Hugging Face’s open-source LeRobot library — a camera-equipped arm, a learning framework and public datasets as one set, for hobby money.

At the other end sit the humanoids. The GR00T reference humanoid ships through Unitree late in 2026 — price not yet announced.

3. Sensors — the bottleneck the papers whisper about

Cameras and IMUs are cheap now. Touch is not. Note that the GR00T reference humanoid went out of its way to mount Sharpa Wave’s tactile five-fingered hands — because vision alone cannot tell you “I am holding it, and it is slipping.”

Who verifiably uses this stack

Only cases confirmed by official announcements or engineering blogs:

• Boston Dynamics — trained Spot’s locomotion policy in Isaac Lab and deployed it to the real robot; the pipeline is public on NVIDIA’s technical blog.
• ETH Zurich, the University of Toronto and the AI Institute — co-developed Orbit with NVIDIA, which became today’s Isaac Lab.
• Google DeepMind — maintains MuJoCo and publishes MJX and Playground.
• Allen Institute for AI, Stanford Robotics Center, UC San Diego ARCLab — named first partners for the GR00T reference humanoid.

Meanwhile, the money is betting on the bodies

If the software is free, where does American capital go? Into the machines that will run it.

• Figure AI closed a Series C of more than $1 billion in late 2025 at a valuation reported around $39 billion — total funding above $1.9 billion, with NVIDIA, Intel Capital, LG and Qualcomm Ventures on the list.
• Apptronik added a $520 million extension in February 2026, taking its Series A past $935 million and its valuation to roughly $5.3 billion — backed by Google, Mercedes-Benz and John Deere.

Three 2026 papers, read past the abstract

1. The benchmark said 95%. The camera said under 30%.

LIBERO-Plus (arXiv:2510.13626) perturbed VLA models seven ways — object placement, camera viewpoint, robot pose, language, lighting, background, sensor noise. Moderate changes to camera viewpoint and initial pose alone dropped success from 95% to below 30%. The stranger finding: changing the language barely moved results — because, follow-up experiments showed, the models largely ignore the instruction. They were not understanding the words. They were not listening.

2. A VLA that fits on a consumer GPU

Rethinking the Practicality of VLA Models (arXiv:2602.22663) starts from a practical complaint: current VLAs are too big, too expensive to pretrain, and break when the robot changes. The team built a new benchmark (CEBench: 14,400 simulated plus 1,600 expert real-world trajectories) and proposes LLaVA-VLA: a compact backbone with multi-view perception and action chunking, trained by two-stage post-training instead of costly pretraining — and claims the first end-to-end VLA for mobile manipulation, navigation and grasping in one action space.

3. Zero real demos — and a 35% success rate

Efficient Sim-to-Real Transfer of World-Action Models (arXiv:2606.31101) asks the question our Foundations article ends on: can a model trained purely on synthetic data drive a real robot zero-shot? Built on Cosmos Policy with about 800 synthetic demos per task and zero real-robot demonstrations, the transferred policies hit 35% average success on a real Franka arm across three tasks. The authors call it the first successful sim-to-real transfer of a world-action model.

The four-step path, if you start tonight

• Step 1 — pick a simulator ($0). RTX GPU → Isaac Lab. No GPU → MuJoCo: pip install mujoco, runs on CPU. First milestone: make a hand-built robot model that does not fall over.
• Step 2 — borrow data, do not collect it. Open X-Embodiment already holds 1M+ real trajectories across 22 robot types. In the RT-1-X experiments, low-data domains gained about 50% from other robots’ experience versus training alone.
• Step 3 — do not build the brain. GR00T N1.7 is Apache 2.0. Fine-tune it; do not attempt pretraining — the “unaffordable pretraining cost” paper 2 complains about is exactly that trap.
• Step 4 — hardware last, and small. A $249 Orin Nano Super plus a $130–240 SO-101 arm keeps real-world validation under $500. Buy the $3,499 Thor only after your policy is stable in simulation.

The honest limits

• “Free” is narrow. The licence is $0; the GPUs, the electricity and the data-collection labor are not. And Isaac locks you into NVIDIA RTX hardware.
• Marketing numbers are scene-specific. Genesis’s 430,000× real-time figure is 43 million FPS on one Franka arm over a flat plane. Add water, cloth or many contacts and it will not survive.
• Maturity is research-grade. A 35% zero-shot success rate is an academic milestone — and a machine that fails two times in three on a factory floor.
• The camera warning travels. After LIBERO-Plus, watch any humanoid demo video with one question first: is the camera bolted down?

Frequently asked questions

Is robot simulation software really free?

All six tools in this guide are. MuJoCo is Apache 2.0, Isaac Sim, Isaac Lab and Genesis are open source, Gazebo has been open source for two decades, and the GR00T N1.7 foundation model is Apache 2.0 — which permits commercial products. The catch is hardware: the Isaac family effectively requires an NVIDIA RTX GPU, and training at scale costs real compute and electricity.

What is the difference between Isaac Sim and Isaac Lab?

Isaac Sim is the simulator itself — GPU physics plus photoreal rendering. Isaac Lab is a robot-learning layer that sits on top: it builds reinforcement- and imitation-learning environments and runs thousands of them in parallel on the GPU. It succeeded Isaac Gym in 2024, and its official paper (arXiv:2511.04831) appeared in November 2025 with 105 NVIDIA-affiliated authors.

How much does the NVIDIA Jetson Thor cost?

The Jetson AGX Thor developer kit lists at $3,499 ($3,429 each for ten or more), verified 12 August 2026. That buys 2,070 FP4 teraflops, 128 GB of memory and a 1 TB SSD, running at 40–130 W. For learning, the $249 Jetson Orin Nano Super is the realistic entry point.

Do I need an NVIDIA GPU to get started?

No. MuJoCo installs with one line — pip install mujoco — and runs on an ordinary CPU, including Apple silicon. You need an RTX GPU only when you move to the Isaac family for camera-based learning, and datacenter GPUs only when you train large policies. Start free, on the laptop you have.

What is a robot foundation model?

A single model that takes camera images and a language instruction and outputs robot actions directly — usually called a VLA (vision-language-action) model. NVIDIA’s GR00T N1 line is the best-known open one; N1.7 is Apache 2.0, so a startup can legally ship it inside a product. Our plain-English explainer covers how these models work.

What can I actually build for under $500?

A complete learn-at-home rig: a $249 Jetson Orin Nano Super for on-device inference, plus the Hugging Face SO-101 arm — about $130 if you source parts and print it yourself, or $220–240 as a dual-arm kit. That pairs with the free LeRobot library and public datasets, so the whole pipeline from teleoperation to a trained policy runs on hobby money.