💾 Checkpoints

Genesis provides state save/load functionality for training resumption and episode resets.

Basic Save/Load

import genesis as gs

scene = gs.Scene()
robot = scene.add_entity(gs.morphs.MJCF(file="franka.xml"))
scene.build()

# Simulate
for _ in range(100):
    scene.step()

# Save checkpoint
scene.save_checkpoint("checkpoint.pkl")

# Load in new scene
scene2 = gs.Scene()
robot2 = scene2.add_entity(gs.morphs.MJCF(file="franka.xml"))
scene2.build()
scene2.load_checkpoint("checkpoint.pkl")

State Objects

# Get current state (in-memory)
state = scene.get_state()

# Reset to initial state
scene.reset()

# Reset to custom state
scene.reset(state=state)

RL Episode Resets

scene.build(n_envs=N)

# Snapshot initial state
init_state = scene.get_state()

for episode in range(num_episodes):
    scene.reset(state=init_state)

    for step in range(episode_length):
        scene.step()
        obs, reward, done = get_observations()

        # Reset environments where episode ended
        if done.any():
            done_envs = torch.where(done)[0].tolist()
            scene.reset(state=init_state, envs_idx=done_envs)

Selective Environment Reset

scene.build(n_envs=16)

# Reset all environments
scene.reset()

# Reset specific environments
scene.reset(envs_idx=[0, 2, 5])

# Reset with custom state for specific envs
scene.reset(state=init_state, envs_idx=[1, 3, 7])

State Contents

The SimState object contains:

Solver	State Variables
Rigid	qpos, dofs_vel, links_pos, links_quat
MPM	pos, vel, C, F, Jp, active
SPH	pos, vel, active
PBD	pos, vel, free
FEM	pos, vel, active

Checkpoint File Format

Checkpoints are pickled dictionaries:

{
    "timestamp": time.time(),
    "step_index": scene.t,
    "arrays": {  # Numpy arrays keyed by solver/field
        "RigidSolver.qpos": np.array(...),
        "MPMSolver.pos": np.array(...),
        ...
    }
}

Serialization for Transfer

# Make state serializable (detach from graph)
state = scene.get_state()
state_serializable = state.serializable()

# Now safe to pickle
import pickle
with open("state.pkl", "wb") as f:
    pickle.dump(state_serializable, f)

Important Notes

• Checkpoints require compatible scene configuration (same entities, solver options)

• 32-bit precision may lose ~2e-6 accuracy between save/load

• Use envs_idx parameter for efficient partial resets

• scene.t stores the simulation step count