Initial commit

.gitignore

@@ -0,0 +1,5 @@
+data
+exp
+results
+.idea
+*__pycache__*

README.md

@@ -0,0 +1,20 @@
+## Overview
+Code of ICML 2023 paper: "Live in the Moment: Learning Dynamics Model Adapted to Evolving Policy"
+
+## Usage
+
+MBPO
+> CUDA_VISIBLE_DEVICES=0 python main_pmac.py --env_name 'Humanoid-v2' --num_epoch 400 --exp_name humanoid_mbpo --seed 6
+
+PDML-MBPO
+> CUDA_VISIBLE_DEVICES=0 python main_pmac_tvd.py --env_name 'Humanoid-v2' --num_epoch 400 --exp_name humanoid_pdml --seed 4 --reweight_model TV --reweight_rollout TV
+
+
+## code structure
+During training, 'exp' folder will created aside 'PDML' folder.
+
+## Dependencies
+MuJoCo 1.5 & MuJoCo 2.0
+
+## Reference
+This code is built on a pytorch implementation MBPO: https://github.com/Xingyu-Lin/mbpo_pytorch

ant.py

@@ -0,0 +1,50 @@
+import numpy as np
+from gym import utils
+from gym.envs.mujoco import mujoco_env
+
+class AntEnv(mujoco_env.MujocoEnv, utils.EzPickle):
+    """
+        External forces (sim.data.cfrc_ext) are removed from the observation.
+        Otherwise identical to Ant-v2 from
+        https://github.com/openai/gym/blob/master/gym/envs/mujoco/ant.py
+    """
+    def __init__(self):
+        mujoco_env.MujocoEnv.__init__(self, 'ant.xml', 5)
+        utils.EzPickle.__init__(self)
+
+    def step(self, a):
+        xposbefore = self.get_body_com("torso")[0]
+        self.do_simulation(a, self.frame_skip)
+        xposafter = self.get_body_com("torso")[0]
+        forward_reward = (xposafter - xposbefore)/self.dt
+        ctrl_cost = .5 * np.square(a).sum()
+        contact_cost = 0.5 * 1e-3 * np.sum(
+            np.square(np.clip(self.sim.data.cfrc_ext, -1, 1)))
+        survive_reward = 1.0
+        reward = forward_reward - ctrl_cost - contact_cost + survive_reward
+        state = self.state_vector()
+        notdone = np.isfinite(state).all() \
+            and state[2] >= 0.2 and state[2] <= 1.0
+        done = not notdone
+        ob = self._get_obs()
+        return ob, reward, done, dict(
+            reward_forward=forward_reward,
+            reward_ctrl=-ctrl_cost,
+            reward_contact=-contact_cost,
+            reward_survive=survive_reward)
+
+    def _get_obs(self):
+        return np.concatenate([
+            self.sim.data.qpos.flat[2:],
+            self.sim.data.qvel.flat,
+            # np.clip(self.sim.data.cfrc_ext, -1, 1).flat,
+        ])
+
+    def reset_model(self):
+        qpos = self.init_qpos + self.np_random.uniform(size=self.model.nq, low=-.1, high=.1)
+        qvel = self.init_qvel + self.np_random.randn(self.model.nv) * .1
+        self.set_state(qpos, qvel)
+        return self._get_obs()
+
+    def viewer_setup(self):
+        self.viewer.cam.distance = self.model.stat.extent * 0.5

humanoid.py

@@ -0,0 +1,58 @@
+import numpy as np
+from gym.envs.mujoco import mujoco_env
+from gym import utils
+
+def mass_center(model, sim):
+    mass = np.expand_dims(model.body_mass, 1)
+    xpos = sim.data.xipos
+    return (np.sum(mass * xpos, 0) / np.sum(mass))[0]
+
+class HumanoidEnv(mujoco_env.MujocoEnv, utils.EzPickle):
+    """
+            COM inertia (cinert), COM velocity (cvel), actuator forces (qfrc_actuator),
+            and external forces (cfrc_ext) are removed from the observation.
+            Otherwise identical to Humanoid-v2 from
+            https://github.com/openai/gym/blob/master/gym/envs/mujoco/humanoid.py
+        """
+    def __init__(self):
+        mujoco_env.MujocoEnv.__init__(self, 'humanoid.xml', 5)
+        utils.EzPickle.__init__(self)
+
+    def _get_obs(self):
+        data = self.sim.data
+        return np.concatenate([data.qpos.flat[2:],
+                               data.qvel.flat,
+                               # data.cinert.flat,
+                               # data.cvel.flat,
+                               # data.qfrc_actuator.flat,
+                               # data.cfrc_ext.flat
+                               ])
+
+    def step(self, a):
+        pos_before = mass_center(self.model, self.sim)
+        self.do_simulation(a, self.frame_skip)
+        pos_after = mass_center(self.model, self.sim)
+        alive_bonus = 5.0
+        data = self.sim.data
+        lin_vel_cost = 1.25 * (pos_after - pos_before) / self.dt
+        quad_ctrl_cost = 0.1 * np.square(data.ctrl).sum()
+        quad_impact_cost = .5e-6 * np.square(data.cfrc_ext).sum()
+        quad_impact_cost = min(quad_impact_cost, 10)
+        reward = lin_vel_cost - quad_ctrl_cost - quad_impact_cost + alive_bonus
+        qpos = self.sim.data.qpos
+        done = bool((qpos[2] < 1.0) or (qpos[2] > 2.0))
+        return self._get_obs(), reward, done, dict(reward_linvel=lin_vel_cost, reward_quadctrl=-quad_ctrl_cost, reward_alive=alive_bonus, reward_impact=-quad_impact_cost)
+
+    def reset_model(self):
+        c = 0.01
+        self.set_state(
+            self.init_qpos + self.np_random.uniform(low=-c, high=c, size=self.model.nq),
+            self.init_qvel + self.np_random.uniform(low=-c, high=c, size=self.model.nv,)
+        )
+        return self._get_obs()
+
+    def viewer_setup(self):
+        self.viewer.cam.trackbodyid = 1
+        self.viewer.cam.distance = self.model.stat.extent * 1.0
+        self.viewer.cam.lookat[2] = 2.0
+        self.viewer.cam.elevation = -20