# Axis Robotics

> Axis Robotics develops simulation-based infrastructure for Physical AI, combining large-scale data collection and augmentation to train robots in realistic conditions, enhancing robotic learning and deployment in real-world environments.

- Canonical URL: https://iq.wiki/wiki/axis-robotics
- Categories: Organizations
- Tags: Base, ResearchLab, DataPlatform
- Created: 2026-07-31T16:21:54.595Z
- Last updated: 2026-07-31T16:39:48.512Z
- Source: IQ.wiki — the world's largest blockchain and crypto encyclopedia (https://iq.wiki)

---

**Axis Robotics** is a robotics company focused on simulation-based infrastructure for Physical AI and robotic learning. Founded in 2025, the company develops systems for collecting, augmenting, and training robotic data across simulated and real-world environments. [\[1\]](#cite-id-8W1HcGYe9NraczN6)  

## Overview

![](https://ipfs.everipedia.org/ipfs/Qmcc8JsJKbKKRKuzQATfh1E3TwaSRKUsWEMNkkxTiv29Pf)

Axis Robotics is a physical AI infrastructure project focused on collecting and processing robot training data through browser-based simulation. Users teleoperate simulated robot arms to perform manipulation tasks, generating trajectories containing robot states, object positions, control actions, and task metadata. The system processes these trajectories and uses an IsaacSim-based backend to create additional photorealistic, domain-randomized training samples. Each accepted trajectory receives a unique Data ID recorded on Base to provide a verifiable provenance record.

The AXIS stack consists of four stages: world generation, behavior collection, data refinement, and model training and deployment. Natural-language instructions are converted into structured tasks and simulated environments using a library of 3D assets and automated scene-generation tools. Collected trajectories are cleaned and filtered before being replayed in simulation to generate expanded datasets, which are then used to train imitation-learning policies and Vision-Language-Action models. The resulting policies are evaluated in simulation and tested on physical robots, with the system currently using the Franka Research 3 for sim-to-real validation. [\[3\]](#cite-id-cBIR3pd0F8tdmfI8) 

## Technology

### RoboVerse

RoboVerse is a simulation framework that provides a common interface across multiple robotics simulators, allowing different backends to be used for different stages of data collection, augmentation, prototyping, and evaluation. AXIS uses MuJoCo compiled to WebAssembly for browser-based data collection, IsaacSim on GPU infrastructure for trajectory replay and photorealistic augmentation, and native MuJoCo for rapid development and evaluation. The system is built on MetaSim, which provides simulator-independent scene descriptions, standardized assets, automatic conversion into native simulator formats, and programmatic scene generation, allowing task definitions, trajectories, and assets to move between supported simulation environments.

RoboVerse v2 extends this framework into a broader data-generation and model-development pipeline focused on variation across tasks, environments, behaviors, and visual conditions. It adds support for humanoid, quadruped, and dexterous robot platforms; standardized interfaces for reinforcement learning, imitation learning, and Vision-Language-Action models; automated scene and asset generation; and a pipeline connecting task creation, teleoperation or policy-based data collection, domain randomization, model training, and real-world deployment. Its evaluation approach emphasizes generalization across simulators, layouts, objects, visual conditions, and physical robots rather than performance on a single fixed benchmark. [\[4\]](#cite-id-AS8Rvb7f4sK4bFz4) 

![](https://ipfs.everipedia.org/ipfs/QmVVZg7kFo5mgy7dSzExgweWY5GBWtTwXFravVvRenuUCP)

### MuJoCo WASM

AXIS uses a browser-based teleoperation system built around MuJoCo compiled to WebAssembly, allowing the full physics simulation to run locally without requiring specialized hardware or software installation. Three.js and WebGL provide real-time 3D rendering, while an embedded inverse kinematics solver converts end-effector inputs such as position, rotation, and gripper state into joint-level commands. The physics, rendering, and user interface operate separately to reduce latency, while trajectory data is buffered during operation and packaged with task metadata into standardized JSON files after an embedded checker verifies task completion.

The system also provides a common control method across different robot embodiments, including parallel-jaw grippers and dexterous hands. Predefined grasp templates represent different manipulation configurations, and operators adjust a single parameter to interpolate between an open rest position and the selected grasp, reducing complex multi-joint hand control to one input axis. New robot embodiments can be incorporated by defining their rest position and relevant grasp templates without changing the underlying teleoperation system. [\[4\]](#cite-id-AS8Rvb7f4sK4bFz4)  

### Realistic Augmentation

AXIS processes browser-collected robot trajectories on Linux GPU servers using IsaacSim. During replay, the system applies photorealistic rendering and domain randomization across lighting, textures, materials, backgrounds, camera settings, and physical parameters such as mass, friction, damping, and joint compliance. These variations are intended to expose models to different visual and physical conditions rather than training them against a single fixed simulation environment. The augmentation system also generates observations from multiple virtual camera viewpoints, producing RGB, depth, and segmentation data from each trajectory. A single original demonstration can therefore be expanded into thousands of samples that vary in appearance, viewpoint, and physical conditions while preserving the original behavior. This approach increases the size and variability of the training dataset while maintaining consistency with the demonstrated task. [\[4\]](#cite-id-AS8Rvb7f4sK4bFz4) 

### Domain Randomization

The RoboVerse Domain Randomization (DR) system is designed to reduce overfitting in robotic reinforcement learning and imitation learning by varying visual and physical conditions within simulated environments. It separates scene lifecycle management from property editing: the SceneRandomizer handles creating, removing, and replacing objects, while dedicated randomizers modify materials, lighting, and camera parameters. An ObjectRegistry provides a common interface for accessing simulation objects, while a three-layer scene hierarchy separates the broader environment, manipulation workspace, and visual distractors. The system supports both physics-enabled static objects and runtime-managed visual objects.

Randomization is applied progressively to support controlled experiments, beginning with a baseline and adding scene geometry and materials, lighting, and camera changes at successive levels. Independent random number generators and fixed seeds allow experiments to be reproduced, including during distributed training, while the system can work across IsaacLab for physics and IsaacSim for rendering. It also incorporates automated asset handling, including cloud-based asset datasets and URDF-to-USD conversion, with memory-efficient scene operations intended for repeated environment resets. [\[4\]](#cite-id-AS8Rvb7f4sK4bFz4) 

### Data Pipeline

The AXIS data pipeline automates task creation before human data collection and trajectory processing afterward. Its Task Generation Engine converts natural-language instructions into structured manipulation tasks, selects or generates 3D assets from a database of more than 2,000 objects, creates spatial layouts, validates the resulting scenes, and generates task-specific success checks. It then samples variations in object positions, clutter, and camera viewpoints to produce different instances of the same task. The platform covers more than 90 task categories, including pick-and-place, pushing, sorting, pouring, opening and closing, stacking, and multi-step manipulation.

After teleoperation, trajectories are filtered for corrupted data, implausible state changes, and failed task conditions, while static frames are removed and remaining movements are smoothed and upsampled from 6–8 Hz to 20 Hz. In tests across 16 LIBERO benchmark tasks, the process reduced mean acceleration by 54% and mean jerk by 47%. Cleaned trajectories are then replayed in IsaacSim, where lighting, materials, textures, physics parameters, backgrounds, and camera viewpoints are varied to generate multi-view RGB, depth, segmentation, and other annotations while preserving the original task behavior. The resulting datasets are structured for use in Vision-Language-Action and imitation-learning training. [\[5\]](#cite-id-rhAWj0Tc9ezYv8VF) 

![](https://ipfs.everipedia.org/ipfs/QmdoG3NChgqxg1QcYgnRM3KpJxr4zqYwuf7etTdzsYb1Yx)

### Training

AXIS trains two main types of robot policies using curated and augmented simulation data: task-specific behavior-cloning and diffusion-policy models, and general-purpose Vision-Language-Action (VLA) models. The training process compares raw teleoperation data, cleaned trajectories, and cleaned trajectories augmented through IsaacSim, using consistent task definitions and training budgets to measure the effect of data processing and augmentation. Policies are evaluated in simulation under varied object configurations, textures, lighting, physics parameters, and scene layouts, then tested for sim-to-real transfer on a Franka Research 3 robotic arm across tasks such as pick-and-place, drawer manipulation, pouring, and button pressing.

The training system is intended to assess whether policies trained primarily on simulated and synthetic data can operate under physical-world conditions. Reported research from the project includes GraspVLA, which examined simulation-based training for robotic grasping, and SkillBlender, which addressed combining robotic skills across different embodiments. The system has several stated limitations, including differences between MuJoCo and IsaacSim during trajectory replay, limited dense labeling and task-success signals in teleoperation data, and the need for embodiment-specific adaptation when transferring policies between robots with substantially different kinematics and control characteristics. [\[6\]](#cite-id-NBgGylZzDCEIUUBk) 

## Architecture

The AXIS architecture is organized into four technical layers: **World Generation, Behavior Collection, Data Refinement, and Model Training & Deployment**. L1 uses natural-language task specifications and an asset library to create simulator-compatible environments through the MetaSim scene description system. L2 allows users to control simulated robots through browser-based interfaces, converting inputs from devices such as keyboards, mobile phones, gamepads, and VR into standardized robot trajectories, with accepted data assigned on-chain provenance records on [Base](https://iq.wiki/wiki/base). L3 cleans these trajectories and uses IsaacSim-based GPU processing and domain randomization to expand demonstrations into larger datasets containing variations in geometry, materials, lighting, camera settings, and physics.

L4 uses the resulting datasets to train behavior-cloning, diffusion-policy, and Vision-Language-Action models, which are evaluated for generalization across different simulators, environments, and robot embodiments before sim-to-real testing on physical hardware. The architecture separates responsibilities between layers and uses standardized data interfaces so individual components can be modified without requiring changes throughout the system. Provenance is maintained across the pipeline, linking trajectories, augmented datasets, and trained models back to their original data sources through Data IDs. [\[3\]](#cite-id-cBIR3pd0F8tdmfI8) 

![](https://ipfs.everipedia.org/ipfs/QmRyxouUPtRRUgrERghoFeJLcNuz9Q4XjDu85Ai4SHUysM)

## Funding

In July 2026, AXIS announced a $12 million seed round led by [Hack VC](https://iq.wiki/wiki/hack-vc), with participation from Nomad Capital, Pi Core Team Ventures, 10K Ventures, and angel investors. The company said the funding would support development of a data infrastructure platform for physical AI that combines large-scale simulation, real-world egocentric data collection, and human-in-the-loop post-training to generate structured robotic datasets. AXIS described the system as a closed-loop data pipeline intended to support the ongoing development and refinement of physical AI models. [\[2\]](#cite-id-hsr3oziFCnD8EdlX) 

## Partnerships

* Dexmal
* Geely
* Stepfun
* Lotus
* Booster
* Manycore