Generate highly complex mechanical SimReady assets with Mad Max

An object’s dimensions and surface geometry affect how a robot interacts with it. A handle needs enough clearance for a gripper. Components in an assembly need to fit together. An articulated mechanism needs to represent the parts that move during a task.

Mad Max improves the representation of these details through more accurate dimensions, curved surfaces, and refined mesh topology. It generates articulated components that allow moving parts to be represented and manipulated individually in simulation.

These capabilities support work on contact-rich manipulation, including assembly and bimanual tasks. In these settings, the asset needs to capture how components move and make contact throughout an interaction.

Mad Max generates moving parts as separate bodies connected by joints, with limits on how far each one travels. A robot can press a key, rotate a cam or open a hinge, and the part moves the way the real mechanism does.

Mechanisms keep their internal components as distinct parts, so the pieces that carry a load or make contact during a task are represented in simulation rather than merged into a single shell.

Mad Max’s modeling improvements work alongside Palatial’s physics capabilities to prepare assets for manipulation. Surfaces carry physically based materials, graspable regions are affordance tagged, and collision geometry is built so that contact resolves cleanly.

Physically based rendering, or PBR, represents how materials respond to lighting. For camera-based robot policies, material appearance is part of the visual input used during training, so PBR shaders give an asset’s surfaces more realistic rendering.

Affordance tagging marks where an object is meant to be grasped, such as a handle, a rim or an edge, so that training can tell the robot where to take hold. A heat map visualizes the resulting grasp predictions over the asset, from lower to higher confidence.

Stable contact normals come from a stable solve. When an asset has correct physical properties, well-built collision proxies, and appropriate simulator settings, the object does not jitter, and the contact normals the solver reports stay consistent and better approximate real-world contact forces. Poor collision geometry or wrong solver parameters cause jitter, and the normals turn noisy with it. Palatial's physics workflow prepares the properties, proxies and settings that produce the stable solve.

Semantic tags provide labels for use in reinforcement learning datasets and downstream data pipelines, so information about an asset travels with its geometry, materials, and physics properties.

Tags follow the part structure. On a pair of headphones, the headband, headband padding, swivel yokes, ear cushions, and earcups are each labeled, so training data can refer to the exact part a robot is holding.

In addition, each part carries a non-visual material tag, so assets can support a broader range of sim sensor suites such as lidar, radar, and ultrasonic.

Mad Max assets are validated in simulation before delivery, so they are checked in a physics engine rather than only in a viewer. Assets are prepared for Isaac Sim, MuJoCo and Newton.

Mad Max is available now in the web app, through the MCP server for coding agents, and in the Isaac Sim plugin, which brings assets straight into an Isaac Sim stage.

Researchers often handle modeling, mesh cleanup, articulation, collision tuning, materials, and annotation as separate preparation steps. Palatial brings that work into a single coordinated asset-generation workflow.

The ambition is straightforward: spend less time preparing the objects in your simulation and more time teaching robots what to do with them.