Isaac Lab
GPU-parallel reinforcement learning
THE STUDIO FOR PHYSICAL AI
Design.
Simulate.
Train. Deploy.
Describe a robot task once in Wano Studio. Wano Framework compiles it to Isaac Lab, MuJoCo, Gazebo and Webots, records LeRobot-ready datasets, and deploys trained policies to real ROS 2 robots. Local-first.
EARLY ACCESS · Q4 2026
Works with the stack you already use
Isaac Lab
MuJoCo
Robot simulation
2
Datasets & trainingA VISUAL JOURNEY / 8 CHAPTERSPHOTO + VIDEO WORKFLOWS · PRODUCT VISION
One small action.
A world of learning.
Follow one cube from a desk photograph to a real robot. See how a scene becomes variations, variations become data, and data becomes a learned skill.
Follow the experienceTHE TASK WE’LL FOLLOW
SAME INTENT.“Pick up the cube and place it in the tray.”
EVERY CHAPTER.
STEP01
WANO / HOW IT LEARNS
Photo to 3D scene
Start with the worldyou already have.
Take a photograph of a desk. AI identifies the objects and their arrangement, then rebuilds the workcell using editable 3D assets from the library. Refine the geometry, physics and task zones in the studio.
IN
Desk photograph
OUT
Editable 3D blueprint
Workflow details
IN
Desk photograph
OUT
Editable 3D blueprint
A table stays a table. The cube, blue cylinder, tray and surrounding space become objects the simulator can work with.
New workflow · product visionILLUSTRATED PIPELINE
SCENE INPUTA photograph becomes a scene.
PHOTO INPUTASSET LIBRARYEDITABLE 3D SCENE
STEP02
WANO / HOW IT LEARNS
Synthetic teleoperation
A human shows it.A robot learns the motion.
A video captures a hand picking up an object and placing it in a tray. Identify the hand, the manipulated object and the action over time. Reconstruct the scene, then retarget the gesture to the robot in simulation.
Téléopération synthétique VIDEO-DRIVEN DEMONSTRATIONS
IN
Human demonstration video
OUT
Robot motion in simulation
Workflow details
IN
Human demonstration video
OUT
Robot motion in simulation
Synthetic teleoperation: demonstrations generated from video, without manually driving the robot for every take.
Téléopération synthétiqueILLUSTRATED PIPELINE
The same object. The same intent. A new embodiment.
HAND + OBJECT TRACKINGACTION TRAJECTORYROBOT REPLAY
STEP03
WANO / HOW IT LEARNS
One scene. Many variations.
Keep the task.Change the world around it.
One demonstration branches into a family of variations. Change the cube’s colour, the tabletop, the wall, the object positions or the motion path. The goal stays the same: put the object in the tray.
IN
One reconstructed demonstration
OUT
Task-preserving variants
Workflow details
IN
One reconstructed demonstration
OUT
Task-preserving variants
Appearance, placement and movement vary together. Every branch adds another way to learn the same skill.
DOMAIN RANDOMIZATIONILLUSTRATED PIPELINE
A family tree of experiences.
SOURCE DEMONSTRATIONAPPEARANCEPLACEMENT + MOTION
STEP04
WANO / HOW IT LEARNS
Parallel simulation
One scene.Many worlds running at once.
Send the scene and its variants to the GPU. Parallel virtual environments replay the behaviour under different conditions. Each environment produces its own observations, actions and outcomes.
IN
Blueprints + GPU compute
OUT
Parallel simulation episodes
Workflow details
IN
Blueprints + GPU compute
OUT
Parallel simulation episodes
The workstation authors the scene. The GPU scales the experience. More environments mean more opportunities to collect useful episodes.
ILLUSTRATED SIMULATIONILLUSTRATED PIPELINE
Compute fans out. Experience multiplies.
GPU COMPUTEPARALLEL ENVIRONMENTSSYNCHRONIZED EPISODES
STEP05
WANO / HOW IT LEARNS
Synthetic datasets
Turn simulated experienceinto training data.
Every accepted episode becomes a synchronized sequence: what the robot saw, what it did, and the instruction it followed. Review the trajectories and keep useful examples before exporting the dataset.
IN
Simulation recordings
OUT
Curated synthetic dataset
Workflow details
IN
Simulation recordings
OUT
Curated synthetic dataset
Images, actions and language stay aligned. Export in the formats your training stack expects, including LeRobot, HDF5 and RLDS.
SYNTHETIC DATA GENERATIONILLUSTRATED PIPELINE
Different streams. One aligned record.
OBSERVATIONSACTIONSLANGUAGE
STEP06
WANO / HOW IT LEARNS
Train the VLA
Give a foundation modelyour robot’s experience.
Bring a pretrained vision-language-action model and the curated dataset together on the GPU. Fine-tuning adapts the model to the task, its objects and the robot’s actions. Follow the run and keep its checkpoints.
IN
Dataset + pretrained VLA
OUT
Task-adapted model weights
Model families & workflow details
IN
Dataset + pretrained VLA
OUT
Task-adapted model weights
Vision describes what the robot sees. Language describes the goal. Action is what the robot does next.
OpenVLAOpen-source VLAπ₀ / π₀.₅Physical IntelligenceIsaac GR00TNVIDIA
Fine-tuning and deployment use model-specific data and robot adapters.
MODEL ADAPTATIONILLUSTRATED PIPELINE
Data changes the weights. The weights change the behaviour.
CURATED DATAGPU FINE-TUNINGADAPTED WEIGHTS
STEP07
WANO / HOW IT LEARNS
Validate the skill
Learn the skill.Not just one arrangement.
Run the trained policy in held-out environments. Change colours, positions and trajectories again, then inspect whether the robot still completes the task. Use the failures to decide what to record or train next.
IN
Adapted policy + unseen scenes
OUT
Evaluated robot behaviour
Workflow details
IN
Adapted policy + unseen scenes
OUT
Evaluated robot behaviour
A policy should handle new conditions, not simply replay a familiar frame. Validation closes the loop before deployment.
GENERALIZATION CHECKILLUSTRATED PIPELINE
New conditions. The original goal.
HELD-OUT SCENESPOLICY EVALUATIONFEEDBACK LOOP
STEP08
WANO / HOW IT LEARNS
Deploy to the real world
The experience was virtual.The action is real.
Export the evaluated policy and connect it to the real robot through the deployment runtime. The robot observes its surroundings, receives the task instruction and predicts the actions that move the object into the tray.
IN
Validated skill + task instruction
OUT
Real-world robot action
Workflow details
IN
Validated skill + task instruction
OUT
Real-world robot action
The learned skill runs on the robot: camera images and a task instruction go in, joint actions come out. The robot repeats this loop as it moves.
SIMULATION → HARDWAREILLUSTRATED PIPELINE
The same task, all the way to the real world.
LEARNED SKILLJOINT ACTIONSREAL-WORLD RESULT
CAPTURESIMULATEGENERATE DATATRAINDEPLOY
01 / THE FRAMEWORK
One declarative blueprint.
Three simulation targets.
Describe your robot, scene, sensors and task once. Wano compiles that same blueprint for Isaac Lab, MuJoCo and Gazebo, and generates a ready-to-run ROS 2 package for real hardware.
wano®RoboticsFRAMEWORKTHE SINGLE SOURCE OF TRUTHYour Wano
Your Wano
blueprint.
One task. One editable description.
- Robot
- SO-ARM100
- Scene
- Table, objects, task zones
- Sensors
- Top + wrist cameras
- Task
- Pick and place
pick_place.blueprint.yaml
MuJoCo
Fast, contact-rich physics
Gazebo
ROS-native robot simulation
AND FOR REAL HARDWARE
2A ready-to-run ROS 2 package.Bring the same robot and task into the real world.
Change the blueprint, not four codebases.
150robots ready to use
MCPClaude Code supported
The blueprint is the editable description Wano uses to generate the simulator-specific files. Change a robot, a sensor or a task in one place.
pick_place.blueprint.yaml YAML
# One source of truth.
task: pick_and_place
robot: so_arm100 # 1 of 150
sensors: [cam_top, cam_wrist]
targets: [isaac_lab, mujoco, gazebo, ros2]Every simulator has its own scene format. Every dataset format needs manual conversion. Sim-to-real is a separate engineering project. A single manipulation task means gluing together Isaac Sim, Python training code, dataset tooling and ROS 2 — by hand, every time.
01
Four scene formats
URDF, MJCF, USD, SDF — each one re-authored by hand.
02
Manual dataset conversion
Every training stack expects its own layout and stats.
03
Sim-to-real as a side project
The ROS 2 package is written twice: once in sim, once for real.
04
Glue code that rots
Isaac Sim + Python + dataset tooling + ROS 2, re-wired per task.
02 / INSIDE WANO STUDIO
Six steps.
One studio.
Every screen below uses the Wano Studio light interface, populated with one coherent robotics workflow.
01SCENE
Build the environment in a 3D editor
Drop a table, props, lights and cameras into the scene, place them with the gizmo, and set the task zones. Everything you place is written into the blueprint — the single declarative source the studio compiles from.
41 ready-made items across 7 categories
02ROBOT
Pick from 150 models, or bring your own
Arms, mobile bases, quadrupeds, humanoids, hands and drones ship with the studio, each with its actuators, sensors and meshes already resolved. Import your own from a URDF or MJCF file when the catalog doesn’t have it.
Cloning a model copies it into your project, fully editable
03TASK
Frame the mission in a few answers
Choose the behavior, robot and environment while the live recap stays in view. Wano checks the composition as you go, then seeds the complete mission — task zones, success condition and training branch included.
One guided flow keeps robot, scene and learning branch consistent
04DEMONSTRATION
Take control and record the behavior
Drive the robot from one cockpit while the 3D workcell, controls, telemetry and latest takes stay visible together. Every accepted motion becomes a synchronized episode ready for review.
One workspace for control, observation and take-by-take recording
05CURATION
Replay every episode before it becomes data
One timeline drives the 3D replay, task phases and synchronized signals. Review the motion, inspect the active steps and accept or reject the take with the evidence still on screen.
The replay, task phases and verdict stay locked to the same frame
06TRAINING
Watch every run converge
Launch a training run from the curated data, then follow convergence and compare every run. Progress, checkpoints and GPU state stay beside the metrics, so the run is readable without leaving the studio.
Convergence, checkpoints and GPU health share one live view
BRING IN
Start with what you have.
Robots and scenes from the formats you already have.
TAKE OUT
Ready for your training stack.
Datasets and policies in the formats training stacks expect.
DEPLOY
Into the real world.
Trained policies talking to real hardware.
03 / MEET YOUR NEXT ROBOT
Explore all 150 Different bodies.
One language.
Arms, quadrupeds, humanoids,
hands and mobile bases.
01
SO-ARM100
arm · 6 DoF02
ALOHA 2
bimanual03
Unitree G1
humanoid04
Spot
quadruped05
Unitree Go2
quadruped06
uFactory Lite 6
arm · 6 DoF07
xArm 5
arm · 5 DoF08
LEAP Hand
hand · 16 DoF09
Allegro Hand
hand · 16 DoF10
TurtleBot 3
mobile base11
Fourier GR-1
humanoid12
AgileX Piper
arm · 6 DoF150 models, ready to use — or import your own URDF or MJCF.
04 / ENGINEERED DIFFERENTLY
Compile, don’t port.
Local, not rented.
Local-first
Your models, your GPUs, your data. Everything runs on your machine — nothing leaves it. The architecture is already split for a SaaS tier, but local is the product, not the demo.
Agent-native
39 MCP tools expose the whole studio to AI agents. A Guard Engine validates every agent action before it touches your scene, your dataset or your robot.
One blueprint, six targets
Isaac Lab, MuJoCo, Gazebo, Webots, plus MuJoCo WASM and Rapier in the browser — from one declarative source. Cross-validate physics instead of trusting one engine.
Proven end-to-end
196 test files. Real torch networks driving simulated robots. Real ROS 2 Humble driven by generated code. Every claim on this page is a passing test, not a render.
BUILT TO SCALE WITH YOU
Local today.
Cloud power the day you need it.
Wano ships local-first: your machine, your GPUs, your data. When a run outgrows the workstation — a hundred parallel environments, a full fine-tune — the studio connects to managed GPUs from the same window. Same blueprint, same datasets, nothing to re-learn.
PLANNED AFTER THE LOCAL RELEASE
01
Your machine
Author, simulate and record locally. That never stops working, with or without an account.
02
One click
Point the run at managed GPUs. No cluster to provision, no YAML, no DevOps on your side.
03
Same pipeline
Datasets and checkpoints come back where you left them, ready to deploy to the robot.
Opt-in by design: nothing leaves your machine unless you send it.
05 / ROADMAP 2026
From here.
To what’s next.
Shipped, in progress, and what comes next.
Two quarters already in your hands; the third is landing now.
Q12026
ShippedWeb studio · 3D scene editor · Isaac Sim streaming
Q22026
ShippedWano Framework: blueprint → multi-sim · MCP server + Guard Engine · 150-robot catalog
Q32026
In progressMulti-sim Docker E2E · multi-robot instancing · domain randomization · LeRobot / GR00T / HDF5 / RLDS datasets · episode viewer · GPU fine-tuning
Q42026
PlannedEarly access · VR teleoperation (Quest 3, CloudXR) · SaaS tier
06 / THE DETAILS
Good questions.
Straight answers.
Questions engineers actually ask.
No. MuJoCo, Gazebo and Webots are first-class targets, and two physics engines run directly in the browser (MuJoCo WASM, Rapier). Isaac Lab is one of six targets — the one you’ll want for GPU-parallel RL.
No. Scenes, datasets, model weights and training runs stay local. A SaaS tier is planned for teams that want managed GPUs — it will be opt-in, not the default.
150 models ship in the catalog — arms (SO-ARM100, ALOHA 2), quadrupeds (Spot), humanoids and more. Anything with a URDF or MJCF can be imported.
Yes. Generated Python, ROS 2 packages and datasets are plain files on your disk, in standard formats (LeRobot, HDF5, RLDS). No lock-in by construction: if you stop using Wano, the code still runs.
Q4 2026, in small cohorts. ROS 2 developers, Isaac Lab users and LeRobot teams get priority. Register below and we’ll be in touch.
WAITLIST OPEN · Q4 2026
Build your first robot policy this year.
Small cohorts. Priority to ROS 2, Isaac Lab and LeRobot teams.
Subscribing to the WanoRobotics newsletter.