SO101-Nexus
Concepts

Backend Support

Simulator backends, environment availability, and key differences.

SO101-Nexus supports MuJoCo and ManiSkill as simulation backends. Genesis is on the roadmap but not yet implemented.

Task Availability

All five tasks are available on both backends:

TaskMuJoCoManiSkill
PickLiftSO-101SO-100, SO-101
PickAndPlaceSO-101SO-100, SO-101
ReachSO-101SO-100, SO-101
LookAtSO-101SO-100, SO-101
MoveSO-101SO-100, SO-101

Vectorization

ManiSkill runs multiple simulation instances in a single GPU-accelerated process by passing num_envs directly to gym.make(). MuJoCo environments must be vectorized through Gymnasium's standard gym.vector.SyncVectorEnv wrapper.

Key Differences

  • Robot models -- MuJoCo environments use the SO-101 arm. ManiSkill has separate SO-100 and SO-101 variants for each task.
  • Object support -- MuJoCo supports CubeObject, YCBObject, and MeshObject. ManiSkill supports CubeObject and YCBObject only.
  • Observations -- Both backends use the same composable observation components (observations parameter). See Observations for details.
  • GPU batching -- ManiSkill supports num_envs for GPU-accelerated parallel simulation. MuJoCo uses CPU-based vectorization.

For details on object types, see the Objects API reference. For observation configuration, see Observations.

On this page