use super::config::HumanoidStandupConfig;
use crate::env::environment::{Environment, Experience, Terminal};
use crate::env::{environment::Error, mujoco::mjenv::MjEnv};
use nalgebra::DVector;
use rand::{rngs::StdRng, Rng, SeedableRng};
use std::collections::HashMap;
pub struct MujocoHumanoidStandupEnv {
pub env: MjEnv,
pub config: HumanoidStandupConfig,
pub init_qpos: Vec<f64>,
pub init_qvel: Vec<f64>,
steps: usize,
}
impl MujocoHumanoidStandupEnv {
pub fn new(config: Option<HumanoidStandupConfig>) -> Result<Self, Error> {
let config = config.unwrap_or_default();
let env = MjEnv::new(&config.xml_file, config.frame_skip)?;
Ok(Self {
init_qpos: env.init_qpos().into(),
init_qvel: env.init_qvel().into(),
config,
env,
steps: 0,
})
}
fn _get_obs(&self) -> Result<DVector<f64>, Error> {
let mut observation = Vec::new();
let mut position = self.env.qpos().to_vec();
let velocity = self.env.qvel().to_vec();
let com_inertia = if self.config.include_cinert_in_observation {
let cinert = self.env.cinert();
let mut cinert_flat = Vec::new();
for body_cinert in cinert.iter().skip(1) {
for &val in body_cinert.iter() {
cinert_flat.push(val);
}
}
cinert_flat
} else {
Vec::new()
};
let com_velocity = if self.config.include_cvel_in_observation {
let cvel = self.env.cvel();
let mut cvel_flat = Vec::new();
for body_cvel in cvel.iter().skip(1) {
for &val in body_cvel.iter() {
cvel_flat.push(val);
}
}
cvel_flat
} else {
Vec::new()
};
let actuator_forces = if self.config.include_qfrc_actuator_in_observation {
let qfrc = self.env.qfrc_actuator();
qfrc[6..].to_vec() } else {
Vec::new()
};
let external_contact_forces = if self.config.include_cfrc_ext_in_observation {
let cfrc_ext = self.env.cfrc_ext();
let mut cfrc_ext_flat = Vec::new();
for body_cfrc in cfrc_ext.iter().skip(1) {
for &val in body_cfrc.iter() {
cfrc_ext_flat.push(val);
}
}
cfrc_ext_flat
} else {
Vec::new()
};
if self.config.exclude_current_positions_from_observation {
position = position[2..].to_vec();
}
observation.extend_from_slice(&position);
observation.extend_from_slice(&velocity);
observation.extend_from_slice(&com_inertia);
observation.extend_from_slice(&com_velocity);
observation.extend_from_slice(&actuator_forces);
observation.extend_from_slice(&external_contact_forces);
Ok(DVector::from_vec(observation))
}
fn _calculate_reward(
&self,
action: &DVector<f64>,
) -> Result<(f64, HashMap<String, f64>), Error> {
let upright_reward = self.config.uph_cost_weight * self.env.qpos()[2] / self.env.timestep();
let ctrl_cost = self._control_cost(action)?;
let impact_cost = self._impact_cost()?;
let total_reward = upright_reward - ctrl_cost - impact_cost + 1.0;
let mut reward_info = HashMap::new();
reward_info.insert("reward_upright".to_string(), upright_reward);
reward_info.insert("reward_ctrl".to_string(), -ctrl_cost);
reward_info.insert("reward_impact".to_string(), -impact_cost);
Ok((total_reward, reward_info))
}
fn _control_cost(&self, _action: &DVector<f64>) -> Result<f64, Error> {
let squared_ctrl: f64 = self.env.ctrl().iter().map(|x| x * x).sum();
Ok(self.config.ctrl_cost_weight * squared_ctrl)
}
fn _impact_cost(&self) -> Result<f64, Error> {
let contact_forces = self.env.cfrc_ext();
let mut total_force = 0.0;
for body_force in contact_forces.iter() {
for &f in body_force.iter() {
total_force += f * f;
}
}
let impact_cost = self.config.impact_cost_weight * total_force;
let clamped_cost = impact_cost.clamp(
self.config.impact_cost_range.0,
self.config.impact_cost_range.1,
);
Ok(clamped_cost)
}
fn _get_reset_info(&self) -> Result<HashMap<String, f64>, Error> {
let mut info = HashMap::new();
info.insert("x_position".to_string(), self.env.qpos()[0]);
info.insert("y_position".to_string(), self.env.qpos()[1]);
info.insert(
"tendon_length".to_string(),
self.env.ten_length().iter().sum(),
); info.insert(
"tendon_velocity".to_string(),
self.env.ten_velocity().iter().sum(),
); info.insert(
"distance_from_origin".to_string(),
(self.env.qpos()[0].powi(2) + self.env.qpos()[1].powi(2)).sqrt(),
);
Ok(info)
}
}
type Action = DVector<f64>;
type State = DVector<f64>;
type Info = HashMap<String, f64>;
impl Environment for MujocoHumanoidStandupEnv {
type Action = Action;
type State = State;
type Info = Info;
fn reset(&mut self, seed: Option<u64>) -> Result<(Self::State, Self::Info), Error> {
self.steps = 0;
self.env.reset_to_initial()?;
let mut rng = match seed {
Some(s) => StdRng::seed_from_u64(s),
None => StdRng::from_os_rng(),
};
let noise_low = -self.config.reset_noise_scale;
let noise_high = self.config.reset_noise_scale;
let mut qpos = self.init_qpos.clone();
if noise_high > noise_low {
for i in 0..qpos.len() {
qpos[i] += rng.random_range(noise_low..noise_high);
}
}
let mut qvel = self.init_qvel.clone();
if noise_high > noise_low {
for i in 0..qvel.len() {
qvel[i] += rng.random_range(noise_low..noise_high);
}
}
self.env.set_state(&qpos, &qvel)?;
let observation = self._get_obs()?;
let info = self._get_reset_info()?;
Ok((observation, info))
}
fn step(
&mut self,
action: Self::Action,
) -> Result<Experience<Self::State, Self::Info, Self::Action>, Error> {
let curr_state = self.state()?;
self.env.do_simulation(action.as_slice())?;
self.steps += 1;
let next_state = self._get_obs()?;
let (reward, reward_info) = self._calculate_reward(&action)?;
let terminated = false;
let truncated = self.steps >= self.config.max_episode_steps;
let mut info = HashMap::new();
info.insert("x_position".to_string(), self.env.qpos()[0]);
info.insert("y_position".to_string(), self.env.qpos()[1]);
info.insert(
"tendon_length".to_string(),
self.env.ten_length().iter().sum(),
); info.insert(
"tendon_velocity".to_string(),
self.env.ten_velocity().iter().sum(),
); info.insert(
"distance_from_origin".to_string(),
(self.env.qpos()[0].powi(2) + self.env.qpos()[1].powi(2)).sqrt(),
);
info.extend(reward_info);
let terminal = Terminal::from_flags(terminated, truncated);
Ok(Experience::new(
curr_state,
reward,
action,
next_state,
info,
terminal,
self.steps as u32,
))
}
fn is_terminal(&self) -> Result<bool, Error> {
Ok(false)
}
fn is_truncated(&self) -> bool {
self.steps >= self.config.max_episode_steps
}
fn state(&self) -> Result<Self::State, Error> {
self._get_obs()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_reset_with_zero_noise_scale_does_not_panic() {
let config = HumanoidStandupConfig {
reset_noise_scale: 0.0,
..HumanoidStandupConfig::default()
};
let mut env = MujocoHumanoidStandupEnv::new(Some(config)).unwrap();
assert!(env.reset(None).is_ok());
}
#[test]
fn test_truncation_at_max_episode_steps() {
let config = HumanoidStandupConfig {
max_episode_steps: 5,
..HumanoidStandupConfig::default()
};
let mut env = MujocoHumanoidStandupEnv::new(Some(config)).unwrap();
env.reset(None).unwrap();
let action = DVector::zeros(env.env.nu());
for _ in 0..4 {
let exp = env.step(action.clone()).unwrap();
assert!(!exp.terminal.is_truncated());
}
let exp = env.step(action).unwrap();
assert!(exp.terminal.is_truncated());
assert!(env.is_truncated());
}
}