common

Module Contents

Classes

PolicySpec
FastNumpyNetwork	Three-layer MLP with tanh activations for deterministic replay.

Functions

num_network_params(→ int)
resolve_site_id(→ int)
get_actuated_joint_ids(→ list[int])
get_joint_state(→ tuple[numpy.ndarray, numpy.ndarray])
set_target_position(→ None)
build_observation(→ numpy.ndarray)
apply_position_delta_control(→ None)
build_model(→ tuple[mujoco.MjModel, mujoco.MjData, ...)

Attributes

NUM_JOINTS
NUM_TUBES
TUBE_MIN
TUBE_MAX
DEFAULT_TARGET
DEFAULT_SIM_STEPS
DEFAULT_CTRL_FREQ
DEFAULT_TOUCH_THRESHOLD
DEFAULT_ACTION_SCALE
DEFAULT_MAX_DELTA

NUM_JOINTS = 6

NUM_TUBES = 5

TUBE_MIN = 0.1

TUBE_MAX = 1.0

DEFAULT_TARGET

DEFAULT_SIM_STEPS = 3500

DEFAULT_CTRL_FREQ = 20

DEFAULT_TOUCH_THRESHOLD = 0.01

DEFAULT_ACTION_SCALE = 0.25

DEFAULT_MAX_DELTA = 0.12

class PolicySpec

input_size: int = 15

hidden_size: int = 32

output_size: int = 6

action_scale: float = 0.25

max_delta: float = 0.12

class FastNumpyNetwork(input_size: int, hidden_size: int, output_size: int, weights: numpy.ndarray)

Three-layer MLP with tanh activations for deterministic replay.

w1

b1

w2

b2

w3

b3

forward(x: numpy.ndarray) → numpy.ndarray

num_network_params(spec: PolicySpec) → int

resolve_site_id(model: mujoco.MjModel, base_name: str) → int

get_actuated_joint_ids(model: mujoco.MjModel, count: int = NUM_JOINTS) → list[int]

get_joint_state(model: mujoco.MjModel, data: mujoco.MjData, joint_ids: list[int]) → tuple[numpy.ndarray, numpy.ndarray]

set_target_position(model: mujoco.MjModel, data: mujoco.MjData, target_site_id: int, target: numpy.ndarray) → None

build_observation(model: mujoco.MjModel, data: mujoco.MjData, joint_ids: list[int], tcp_sid: int, tgt_sid: int) → numpy.ndarray

apply_position_delta_control(model: mujoco.MjModel, data: mujoco.MjData, joint_ids: list[int], action: numpy.ndarray, action_scale: float, max_delta: float) → None

build_model(tube_lengths: numpy.ndarray) → tuple[mujoco.MjModel, mujoco.MjData, int, int, list[int]]