Shared model with sensor fusion for the NavRL task.
Classes:
| Name | Description |
|---|---|
SensorFusionMLPModel |
Shared actor-critic model with sensor fusion. |
SensorFusionMLPModel(obs, obs_groups, obs_set, output_dim, hidden_dims=(256, 256), activation='lrelu', obs_normalization=False, distribution_cfg=None, cnn_cfg=None, cnns: ModuleDict | dict[str, Module] | None = None, lidar_out_dim: int = 128, dyn_obs_out_dim: int = 64)
Bases: CNNModel
Shared actor-critic model with sensor fusion.
Extends :class:CNNModel so the LiDAR branch flows through rsl_rl's
standard CNN mechanism, enabling as_jit / as_onnx export.
The dynamic-obstacle branch uses a custom MLP encoder whose latent is
concatenated with the CNN and 1-D latents before the head MLP.
Initialize encoders and policy/value heads.
Methods:
| Name | Description |
|---|---|
get_latent |
Concatenate 1-D, CNN, and dynamic-obstacle latents. |
as_jit |
Return a TorchScript-compatible wrapper. |
as_onnx |
Return an ONNX-compatible wrapper. |
get_latent(obs, masks=None, hidden_state=None) -> torch.Tensor
Concatenate 1-D, CNN, and dynamic-obstacle latents.
as_jit() -> nn.Module
Return a TorchScript-compatible wrapper.
as_onnx(verbose: bool = False) -> nn.Module
Return an ONNX-compatible wrapper.