newton.actuators.ControllerNeuralMLP

class newton.actuators.ControllerNeuralMLP(model_path)

Bases: Controller

MLP-based neural network controller.

Uses a pre-trained MLP to compute joint effort from position error and velocity error history.

The network receives concatenated, scaled position-error and velocity-error history as input. The output is multiplied by effort_scale to convert from network units to physical effort [N or N·m]. All three scale factors default to 1.0 (no scaling).

Configuration parameters (input_order, input_idx, pos_scale, vel_scale, effort_scale) are read from the checkpoint metadata, falling back to defaults when absent. Supported checkpoint formats: TorchScript (.pt saved with torch.jit.save) and state-dict bundles ({"model": state_dict, "metadata": {...}} saved with torch.save).

classmethod resolve_arguments(args)

__init__(model_path)

Initialize MLP controller from a checkpoint file.

Supported checkpoint formats: TorchScript (.pt saved with torch.jit.save) and state-dict bundles ({"model": state_dict, "metadata": {...}} saved with torch.save).

Configuration is read from checkpoint metadata:

• input_order (str): "pos_vel" or "vel_pos" (default "pos_vel").

• input_idx (list[int]): history timestep indices (default [0]).

• pos_scale (float): position-error scaling (default 1.0).

• vel_scale (float): velocity-error scaling (default 1.0).

• effort_scale (float): output effort scaling (default 1.0).

Parameters:

model_path (str) – Path to the checkpoint (.pt).

compute(positions, velocities, target_pos, target_vel, feedforward, pos_indices, vel_indices, target_pos_indices, target_vel_indices, forces, state, dt, device=None)

finalize(device, num_actuators)

is_graphable()

is_stateful()

state(num_actuators, device)

update_state(current_state, next_state)

SHARED_PARAMS: ClassVar[set[str]] = {'model_path'}