BipedalLocomotion::ML::MANNAutoregressive class
#include <BipedalLocomotion/ML/MANNAutoregressive.h>

MANNAutoregressive is a class that allows to perform autoregressive inference with Mode-Adaptive Neural Networks (MANN).

MANN Autoregressive

The following diagram shows how the MANN network is used inside the MANNAutoregressive class. The output of MANN is given to two blocks, one computes the kinematically feasible base position, while the other blends it to te input provided by the user

Base classes

template<class _Input, class _Output> class BipedalLocomotion::System::Advanceable<MANNAutoregressiveInput, MANNAutoregressiveOutput>: Basic class that represents a discrete system.

Public types

struct AutoregressiveState: AutoregressiveState contains all quantities required to reset the Advanceable The base position trajectory, facing direction trajectory and base velocity trajectories are written in a bidimensional local reference frame L in which we assume all the quantities related to the ground-projected base trajectory in xi and yi to be expressed.
enum class SupportFoot { Left, Right, Unknown }: SupportFoot is an enum that contains the support foot considered by the MANN network.

Constructors, destructors, conversion operators

MANNAutoregressive(): Constructor.
~MANNAutoregressive(): Destructor.

Public functions

auto setRobotModel(const iDynTree::Model& model) -> bool: Set the robot model.
auto initialize(std::weak_ptr<const ParametersHandler::IParametersHandler> paramHandler) -> bool override: Initialize the network.
auto setInput(const Input& input) -> bool override: Set the input to the autoregressive model.
auto advance() -> bool override: Perform one step of the autoregressive model.
auto reset(Eigen::Ref<const Eigen::VectorXd> jointPositions, const manif::SE3d& basePose) -> bool: Reset the autoregressive model.
auto reset(const AutoregressiveState& autoregressiveState) -> bool: Reset the autoregressive model.
auto isOutputValid() const -> bool override: Check if the output is valid.
auto getOutput() const -> const Output& override: Get the output from MANNAutoregressive.
auto getMANNInput() const -> const MANNInput&: Get the structure that has been used as input for MANN.
auto getAutoregressiveState() const -> const AutoregressiveState&: Get the autoregressive state required to rest MANNAutoregressive.
auto populateInitialAutoregressiveState(Eigen::Ref<const Eigen::VectorXd> jointPositions, const manif::SE3d& basePose, MANNAutoregressive::AutoregressiveState& state) -> bool: Get the autoregressive state required to reset MANNAutoregressive.

Enum documentation

enum class BipedalLocomotion::ML::MANNAutoregressive::SupportFoot

SupportFoot is an enum that contains the support foot considered by the MANN network.

Enumerators
Left	Left foot.
Right	Right foot.
Unknown	This is useful to detect unexpected behaviour.

Function documentation

bool BipedalLocomotion::ML::MANNAutoregressive::setRobotModel(const iDynTree::Model& model)

Set the robot model.

Parameters
model	model of the robot considered by the network. Please load the very same model with the same joint serialization used to train the MANN network.
Returns	true in case of success, false otherwise.

bool BipedalLocomotion::ML::MANNAutoregressive::initialize(std::weak_ptr<const ParametersHandler::IParametersHandler> paramHandler) override

Initialize the network.

Parameters
paramHandler	pointer to the parameters handler.
Returns	True in case of success, false otherwise.

Note

the following parameters are required by the class

Parameter Name	Type	Description	Mandatory
`sampling_time`	`double`	Sampling considered in the inference.	Yes
`root_link_frame_name`	`string`	Name of of the root link frame in the model.	Yes
`chest_link_frame_name`	`string`	Name of of the chest link frame in the model.	Yes
`right_foot_frame_name`	`string`	Name of of the right foot frame in the model.	Yes
`left_foot_frame_name`	`string`	Name of of the left foot frame in the model.	Yes
`forward_direction`	`string`	String containing 'x', 'y' or 'z' representing the foot link forward axis. Currently, only 'x' is supported.	Yes
`mocap_frame_rate`	`int`	Frame rate of the mocap data.	Yes
`past_projected_horizon`	`double`	Number of seconds of the past base horizon considered in the neural network.	Yes

It is also required to define two groups LEFT_FOOT and RIGHT_FOOT that contains the following parameter

Parameter Name	Type	Description	Mandatory
`number_of_corners`	`int`	Number of corners associated to the foot	Yes
`corner_<i>`	`vector<double>`	Position of the corner expressed in the foot frame. It must be from 0 to number_of_corners - 1.	Yes
`on_threshold`	`double`	Distance between the foot and the ground used as on threshold of the trigger to activate the contact. It must be positive.	Yes
`off_threshold`	`double`	Distance between the foot and the ground used as off threshold of the trigger to deactivate the contact. It must be positive.	Yes
`switch_on_after`	`double`	Seconds to wait for before switching to activate from deactivate contact. Ensure it's greater than sampling time.	Yes
`switch_off_after`	`double`	Seconds to wait for before switching to deactivate from activate contact. Ensure it's greater than sampling time.	Yes

Finally it also required to define a group named MANN that contains the following parameter

Parameter Name	Type	Description	Mandatory
`onnx_model_path`	`string`	Path to the `onnx` model that will be loaded to perform inference.	Yes
`projected_base_datapoints`	`int`	Number of samples of the base horizon considered in the model (It must be an even number).	Yes

bool BipedalLocomotion::ML::MANNAutoregressive::setInput(const Input& input) override

Set the input to the autoregressive model.

Parameters
input	input to the model
Returns	true in case of success, false otherwise.

bool BipedalLocomotion::ML::MANNAutoregressive::advance() override

Perform one step of the autoregressive model.

Returns	true in case of success, false otherwise.

bool BipedalLocomotion::ML::MANNAutoregressive::reset(Eigen::Ref<const Eigen::VectorXd> jointPositions, const manif::SE3d& basePose)

Reset the autoregressive model.

Parameters
jointPositions	joint position.
basePose	base pose.
Returns	true in case of success, false otherwise.

bool BipedalLocomotion::ML::MANNAutoregressive::reset(const AutoregressiveState& autoregressiveState)

Reset the autoregressive model.

Parameters
autoregressiveState	the autoregressive state at which you want to reset your model.
Returns	true in case of success, false otherwise.

bool BipedalLocomotion::ML::MANNAutoregressive::isOutputValid() const override

Check if the output is valid.

Returns	true if the output is valid, false otherwise.

const Output& BipedalLocomotion::ML::MANNAutoregressive::getOutput() const override

Get the output from MANNAutoregressive.

Returns	the output of the system.

const MANNInput& BipedalLocomotion::ML::MANNAutoregressive::getMANNInput() const

Get the structure that has been used as input for MANN.

Returns	the MANNInput

const AutoregressiveState& BipedalLocomotion::ML::MANNAutoregressive::getAutoregressiveState() const

Get the autoregressive state required to rest MANNAutoregressive.

Returns	the AutoregressiveState

bool BipedalLocomotion::ML::MANNAutoregressive::populateInitialAutoregressiveState(Eigen::Ref<const Eigen::VectorXd> jointPositions, const manif::SE3d& basePose, MANNAutoregressive::AutoregressiveState& state)

Get the autoregressive state required to reset MANNAutoregressive.

Parameters
jointPositions	joint position
basePose	base pose
state	autoregressive state that will be populated
Returns	true in case of success, false otherwise.