Isaac to mujoco loading

sim/deploy/config.py

@@ -57,7 +57,7 @@ class RobotConfig:
     kds: np.ndarray
     tau_limit: np.ndarray
     robot_model_path: str
-    dt: float = 0.001
+    dt: float = 0.0001  # 0.001
     phase_duration: float = 0.64
     duration: float = 10.0
     decimation: int = 10

sim/deploy/policy.py

@@ -35,7 +35,10 @@ def __init__(
         self.hist_obs = deque()
         for _ in range(frame_stack):
             self.hist_obs.append(np.zeros([1, obs_size], dtype=np.double))
-        self.model = torch.jit.load(model_path)
+        # Load the model.
+        # Changed to load onto same device as other tensors
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        self.model = torch.load(model_path, map_location=self.device)
         self.action = np.zeros((num_actions), dtype=np.double)
 
     @abstractmethod
@@ -167,7 +170,7 @@ def next_action(
         for i in range(self.cfg.frame_stack):
             policy_input[0, i * self.cfg.num_single_obs : (i + 1) * self.cfg.num_single_obs] = self.hist_obs[i][0, :]
 
-        self.action[:] = self.model(torch.tensor(policy_input))[0].detach().numpy()
+        self.action[:] = self.model(torch.tensor(policy_input).to(self.device))[0].detach().cpu().numpy()
 
         self.action = np.clip(self.action, -self.cfg.normalization.clip_actions, self.cfg.normalization.clip_actions)
         return self.action

sim/deploy/run.py

@@ -1,4 +1,5 @@
 """Basic sim2sim and sim2real deployment script.
+          <joint name="lfemurrx" pos="0 0 0" axis="1 0 0" range="-0.349066 2.79253" class="stiff_medium"/>
 
 Run example:
     mjpython sim/deploy/run.py --load_model sim/deploy/tests/walking_policy.pt --world MUJOCO
@@ -21,7 +22,7 @@
 
 from sim.deploy.config import RobotConfig
 from sim.env import stompy_mjcf_path
-from sim.stompy.joints import StompyFixed
+from sim.stompy_legs.joints import Stompy as StompyFixed
 
 
 class Worlds(Enum):
@@ -62,11 +63,18 @@ class MujocoWorld(World):
     def __init__(self, cfg: RobotConfig):
         self.cfg = cfg
         self.model = mujoco.MjModel.from_xml_path(str(self.cfg.robot_model_path))
+
         self.model.opt.timestep = self.cfg.dt
+
         # First step
         self.data = mujoco.MjData(self.model)
+        self.data.qpos = self.model.keyframe("default").qpos
+
         mujoco.mj_step(self.model, self.data)
 
+        # Set self.data initial state to the default standing position
+        # self.data.qpos = StompyFixed.default_standing()
+
     def step(
         self,
         tau: np.ndarray = None,
@@ -109,6 +117,7 @@ def get_observation(self) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarra
         return dof_pos, dof_vel, orientation, ang_vel
 
     def simulate(self, policy: SimPolicy) -> None:
+
         with mujoco.viewer.launch_passive(self.model, self.data) as viewer:
             for step in tqdm(range(int(cfg.duration / cfg.dt)), desc="Simulating..."):
                 # Get the world state
@@ -141,7 +150,7 @@ def __init__(self, cfg: RobotConfig):
         delattr(args, "load_model")
 
         # adapt
-        args.task = "legs_ppo"
+        args.task = "only_legs_ppo"
         args.device = "cpu"
         args.physics_engine = gymapi.SIM_PHYSX
         args.num_envs = 1