Training assets for
robotic arms
Manipulation policies need objects with accurate mass, friction, and collision geometry. Rigyd generates physics-enabled assets optimized for grasping, stacking, and assembly training.
The problem
Why existing workflows fall short.
Grasping needs accurate contact physics
Robotic arm manipulation is highly sensitive to object mass, center of mass, friction coefficients, and collision mesh detail. Wrong values mean failed grasps.
Object diversity improves generalization
Training on a small set of objects produces brittle policies. Thousands of varied objects with realistic physics are needed for real-world robustness.
Center of mass matters
Asymmetric objects like tools, bottles, and electronics have off-center mass distributions. Ignoring this produces unrealistic simulation behavior.
How Rigyd helps
AI-native infrastructure that automates the hard parts.
Precise center of mass estimation
Rigyd analyzes object geometry and identified materials to estimate center of mass position — critical for manipulation planning.
Detailed collision meshes
Convex decomposition captures object geometry at the detail level needed for contact-rich manipulation, not just bounding box approximations.
Domain randomization ready
Physics estimates are calibrated within typical domain randomization variance ranges (mass within 15-20%), enabling robust policy training.
better real-world performance with physics-accurate training
faster than manual physics annotation
per object from upload to SimReady
Train better manipulation policies
Upload objects your robot needs to handle and get SimReady assets in minutes.
Starts at $29/month. 30 credits included.
Frequently asked questions
How does Rigyd estimate center of mass for manipulation?
Volume analysis combined with per-region material identification produces center-of-mass positions accurate enough for grasping and manipulation planning — critical for asymmetric objects like tools, bottles, and electronics.
Does Rigyd work for contact-rich tasks like assembly?
Yes. Convex decomposition captures object geometry at manipulation detail level — not just bounding boxes. Friction and restitution are calibrated for contact-rich training scenarios.
Can I train on domain-randomized physics with Rigyd assets?
Yes. Physics estimates are intentionally calibrated within typical DR variance (mass 15-20%, friction 0.1 coefficient). Use them as realistic baselines that DR can vary around for robust policy transfer.
How does Rigyd handle transparent, reflective, or deformable objects?
Glass, polished metal, and other low-information surfaces fall back to category priors (drinkware glass: 2,500 kg/m³, polished aluminum: 2,700 kg/m³). Reflective surfaces don't affect mass estimation but can confuse material classification — override per-asset if needed. Deformable objects (cloth, soft toys) are flagged for review; Rigyd targets rigid-body simulation, and soft-body physics needs simulator-specific authoring downstream.
Does the collision mesh capture delicate grasp features like handles, rims, and tabs?
Yes. V-HACD parameters are tuned per object class — graspable items get 16-32 convex hulls with concavity threshold 0.001, which preserves handles, mug rims, tool grooves, and other functional concavities. The trade-off is per-asset hull count vs simulation speed; manipulation-grade settings are the default because lost concavities cause sim-to-real grasp failures more often than physics performance bottlenecks. For research benchmarks like the ManiSkill or RoboCasa task suites, where graspable diversity is the rate-limiter, the same 16-32 hull profile transfers cleanly — and tighter (8-12 hulls) settings are available when policy training prioritizes simulation speed over per-grasp accuracy.
Related reading
In-depth guides on robotics simulation, OpenUSD, and SimReady assets.
21 MAY 2026
Synthetic data generation for computer vision in robotics
Real-world labeled robotics data is expensive and slow to capture. Synthetic data is fast and unlimited — but only useful if the underlying simulation has correct physics, semantic labels, and domain randomization. Here's the complete pipeline.
18 MAY 2026
Digital twin creation pipeline for manufacturing
A factory digital twin needs every object to behave physically, not just render. This is the end-to-end pipeline: CAD intake, BIM merge, physics layer, semantic labeling, simulation runtime — at the asset volumes (10K+ unique SKUs) real factories actually contain.
13 MAY 2026
How to set up mass, friction, and joint properties for robot training
The three pillars of robot physics setup — mass, friction, joints — determine whether your trained policy transfers to real hardware. Here's the calibration target for each, the schemas, and the common mistakes that quietly break training.
Explore more
SimReady for Isaac Sim
OpenUSD assets validated for NVIDIA Isaac Sim
Models for MuJoCo
Physics-accurate assets for robot learning
Assets for Gazebo
Simulation-ready models for ROS 2
Warehouse Simulation
50,000+ SKUs need physics properties
Humanoid Robots
Whole-body interaction training assets
Sim-to-Real Transfer
Close the gap with accurate physics data