One Skeleton, Three Benchmarks: Refreshing YCB, LIBERO, and RoboCasa on Isaac Lab-Arena
September 29, 2025
Lightwheel
One-line Context
We’re honored to collaborate with NVIDIA on NVIDIA Isaac Lab - Arena, a Policy Evaluation Framework built on NVIDIA Isaac Lab, an open-source framework for robot learning. Beyond training/evaluation, one of Isaac Lab - Arena’s most useful applications is serving as a benchmark skeleton—combining high-quality scenes/assets with modular task/robot/scene interfaces so that benchmarks can be migrated or extended consistently.
Three Benchmarks Refreshed on Isaac Lab - Arena
1) Lightwheel-YCB
Rebuilt the full set of 106 original objects, plus 19 block cubes for curriculum and coverage
Upgraded visual fidelity (meshes, PBR materials) and physics realism (mass, friction, stiffness)
Covered rigid / articulated / deformable objects, with optimized collision bodies for stability and performance
Released in OpenUSD and MJCF formats; behaviors validated via teleoperation
A reliable, simulation-ready asset evaluation suite for manipulation and contact-rich tasks
Smaller sim-to-real gaps through improved visuals and physics
A strong, common asset foundation for multi-task and multi-platform evaluation
2) Lightwheel-LIBERO
Migrated all 130 tasks to NVIDIA Isaac Lab-Arena
Refreshed environments and assets across kitchens, living rooms, studies, coffee tables, and floors
Extended multi-robot support: Piper, X7s, G1
Standardized demonstrations: 50 human demos per task × robot
Purpose-built for multitask, transfer, and lifelong learning—now comparable and reproducible within a unified skeleton
Enables cross-embodiment studies (different robot morphologies) with consistent evaluation
Unified demonstration specs accelerate IL/RL baselines and reproducibility
3) Lightwheel-RoboCasa
Migrated 138 tasks to NVIDIA Isaac Lab-Arena
Refreshed 100 kitchen scenes (10 layouts × 10 styles) with improved lighting/material/physics consistency
Integrated 2,500+ high-quality assets for task randomization and distributional robustness
Extended multi-robot support: G1, X7s, R1 Pro
Standardized demonstrations: 50 human demos per task × robot
Large-scale, realistic household settings with long-horizon activities—ideal for evaluating generalist agents
Rich assets and randomization enable stronger tests of generalization
Smooth path from “realistic home tasks” to standardized evaluation inside one skeleton
Consistent evaluation: YCB, LIBERO, and RoboCasa run within the same backbone—apples-to-apples comparisons across tasks, scenes, and robots
Higher quality & reproducibility: Upgraded assets/physics + standardized pipelines reduce confounds and speed sim-to-real progress
Rapid extension: High-quality assets and modular interfaces make it easy to add new robots/tasks/benchmarks