Fix minor bugs, fix formatting, enforce consistent type hinting

Author: amacati

benchmark/main.py

@@ -21,7 +21,7 @@ def analyze_timings(times: list[float], n_steps: int, n_worlds: int, freq: float
     # Check for significant variance
     if tmax / tmin > 5:
         print("Warning: step time varies by more than 5x. Is JIT compiling during the benchmark?")
-        print(f"Times: max {tmax:.2e}@{idx_tmax}, min {tmin:.2e}@{idx_tmin}")
+        print(f"Times: max {tmax:.2e} @ {idx_tmax}, min {tmin:.2e} @ {idx_tmin}")
 
     # Performance metrics
     n_frames = n_steps * n_worlds  # Number of frames simulated
@@ -43,7 +43,7 @@ def profile_gym_env_step(sim_config: config_dict.ConfigDict, n_steps: int, devic
     device = jax.devices(device)[0]
 
     envs = gymnasium.make_vec(
-        "CrazyflowEnvReachGoal-v0",
+        "DroneReachPos-v0",
         max_episode_steps=200,
         return_datatype="numpy",
         num_envs=sim_config.n_worlds,
@@ -114,10 +114,10 @@ def main():
     sim_config.controller = "emulatefirmware"
     sim_config.device = device
 
-    print("SIM PERFORMANCE")
+    print("Simulator performance")
     profile_step(sim_config, 100, device)
 
-    print("\nGYM ENV PERFORMANCE")
+    print("\nGymnasium environment performance")
     profile_gym_env_step(sim_config, 100, device)
 
 

benchmark/performance.py

@@ -1,3 +1,7 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
 import gymnasium
 import jax
 import numpy as np
@@ -8,6 +12,9 @@
 import crazyflow  # noqa: F401, ensure gymnasium envs are registered
 from crazyflow.sim.core import Sim
 
+if TYPE_CHECKING:
+    from crazyflow.gymnasium_envs import CrazyflowEnvReachGoal
+
 
 def profile_step(sim_config: config_dict.ConfigDict, n_steps: int, device: str):
     sim = Sim(**sim_config)
@@ -40,8 +47,8 @@ def profile_step(sim_config: config_dict.ConfigDict, n_steps: int, device: str):
 def profile_gym_env_step(sim_config: config_dict.ConfigDict, n_steps: int, device: str):
     device = jax.devices(device)[0]
 
-    envs = gymnasium.make_vec(
-        "CrazyflowEnvReachGoal-v0",
+    envs: CrazyflowEnvReachGoal = gymnasium.make_vec(
+        "DroneReachPos-v0",
         max_episode_steps=200,
         return_datatype="numpy",
         num_envs=sim_config.n_worlds,
@@ -50,28 +57,25 @@ def profile_gym_env_step(sim_config: config_dict.ConfigDict, n_steps: int, devic
     )
 
     # Action for going up (in attitude control)
-    action = np.array(
-        [[[-0.3, 0, 0, 0] for _ in range(sim_config.n_drones)] for _ in range(sim_config.n_worlds)],
-        dtype=np.float32,
-    ).reshape(sim_config.n_worlds, -1)
-
-    # step through env once to ensure JIT compilation
-    _, _ = envs.reset_all(seed=42)
-    _, _, _, _, _ = envs.step(action)
-    _, _ = envs.reset_all(seed=42)
-    _, _, _, _, _ = envs.step(action)
-    _, _ = envs.reset_all(seed=42)
-    _, _, _, _, _ = envs.step(action)
-    _, _ = envs.reset_all(seed=42)
-
-    jax.block_until_ready(envs.unwrapped.sim._mjx_data)  # Ensure JIT compiled dynamics
+    action = np.zeros((sim_config.n_worlds, 4), dtype=np.float32)
+    action[..., 0] = -0.3
+
+    # Step through env once to ensure JIT compilation.
+    # TODO: Currently triggering recompiles also after the first full run. Investigate why and fix
+    # envs accordingly.
+    envs.reset_all(seed=42)
+
+    for _ in range(envs.max_episode_steps + 1):  # Ensure all paths have been taken at least once
+        envs.step(action)
+
+    envs.reset_all(seed=42)
 
     profiler = Profiler()
     profiler.start()
 
     for _ in range(n_steps):
         _, _, _, _, _ = envs.step(action)
-        jax.block_until_ready(envs.unwrapped.sim._mjx_data)
+
     profiler.stop()
     renderer = HTMLRenderer()
     renderer.open_in_browser(profiler.last_session)
@@ -89,17 +93,6 @@ def main():
     sim_config.device = device
 
     profile_step(sim_config, 1000, device)
-    # old | new
-    # sys_id + attitude:
-    # 0.61 reset, 0.61 step  |  0.61 reset, 0.61 step
-    # sys_id + state:
-    # 14.53 step, 0.53 reset |  0.75 reset, 0.88 step
-
-    # Analytical + attitude:
-    # 0.75 reset, 9.38 step  |  0.75 reset, 0.80 step
-    # Analytical + state:
-    # 0.75 reset, 15.1 step  |  0.75 reset, 0.82 step
-
     profile_gym_env_step(sim_config, 1000, device)
 
 

crazyflow/__init__.py

@@ -0,0 +1 @@
+import crazyflow.gymnasium_envs  # noqa: F401, ensure gymnasium envs are registered

crazyflow/gymnasium_envs/__init__.py

@@ -1,11 +1,15 @@
 from gymnasium.envs.registration import register
 
+from crazyflow.gymnasium_envs.crazyflow import CrazyflowEnvReachGoal, CrazyflowEnvTargetVelocity
+
+__all__ = ["CrazyflowEnvReachGoal", "CrazyflowEnvTargetVelocity"]
+
 register(
-    id="CrazyflowEnvReachGoal-v0",
-    vector_entry_point="crazyflow.gymnasium_envs:CrazyflowEnvReachGoal",
+    id="DroneReachPos-v0",
+    vector_entry_point="crazyflow.gymnasium_envs.crazyflow:CrazyflowEnvReachGoal",
 )
 
 register(
-    id="CrazyflowEnvTargetVelocity-v0",
-    vector_entry_point="crazyflow.gymnasium_envs:CrazyflowEnvTargetVelocity",
+    id="DroneReachVel-v0",
+    vector_entry_point="crazyflow.gymnasium_envs.crazyflow:CrazyflowEnvTargetVelocity",
 )

crazyflow/gymnasium_envs/crazyflow.py

@@ -1,16 +1,16 @@
 import math
 import warnings
 from functools import partial
-from typing import Dict, Literal, Optional, Tuple
+from typing import Literal
 
 import jax
 import jax.numpy as jnp
-import numpy as np
 from flax.struct import dataclass
 from gymnasium import spaces
 from gymnasium.vector import VectorEnv
 from gymnasium.vector.utils import batch_space
 from jax import Array
+from numpy.typing import NDArray
 
 from crazyflow.control.controller import MAX_THRUST, MIN_THRUST, Control
 from crazyflow.sim.core import Sim
@@ -19,8 +19,8 @@
 
 @dataclass
 class RescaleParams:
-    scale_factor: jnp.ndarray
-    mean: jnp.ndarray
+    scale_factor: Array
+    mean: Array
 
 
 CONTROL_RESCALE_PARAMS = {
@@ -35,6 +35,12 @@ class RescaleParams:
 }
 
 
+@partial(jax.jit, static_argnames=["convert"])
+def maybe_to_numpy(data: Array, convert: bool) -> NDArray | Array:
+    """Converts data to numpy array if convert is True."""
+    return jax.lax.cond(convert, lambda: jax.device_get(data), lambda: data)
+
+
 class CrazyflowBaseEnv(VectorEnv):
     """JAX Gymnasium environment for Crazyflie simulation."""
 
@@ -103,7 +109,7 @@ def __init__(
         )
         self.observation_space = batch_space(self.single_observation_space, self.sim.n_worlds)
 
-    def step(self, action: Array) -> Tuple[Array, Array, Array, Array, Dict]:
+    def step(self, action: Array) -> tuple[Array, Array, Array, Array, dict]:
         assert self.action_space.contains(action), f"{action!r} ({type(action)}) invalid"
         action = jnp.array(action, device=self.device).reshape(
             (self.sim.n_worlds, self.sim.n_drones, -1)
@@ -139,8 +145,8 @@ def step(self, action: Array) -> Tuple[Array, Array, Array, Array, Dict]:
         return (
             self._get_obs(),
             reward,
-            self._maybe_to_numpy(terminated),
-            self._maybe_to_numpy(truncated),
+            maybe_to_numpy(terminated, self.return_datatype == "numpy"),
+            maybe_to_numpy(truncated, self.return_datatype == "numpy"),
             {},
         )
 
@@ -165,7 +171,7 @@ def _rescale_action(action: Array, control_type: str) -> Array:
         return action * params.scale_factor + params.mean
 
     def reset_all(
-        self, *, seed: Optional[int] = None, options: Optional[dict] = None
+        self, *, seed: int | None = None, options: dict | None = None
     ) -> tuple[dict[str, Array], dict]:
         super().reset(seed=seed)
 
@@ -226,42 +232,37 @@ def _reward() -> None:
 
     @staticmethod
     @jax.jit
-    def _terminated(dones: jax.Array, states: SimState, contacts: jax.Array) -> jnp.ndarray:
+    def _terminated(dones: Array, states: SimState, contacts: Array) -> Array:
         contact = jnp.any(contacts, axis=1)
         z_coords = states.pos[..., 2]
-        below_ground = jnp.any(
-            z_coords < -0.1, axis=1
-        )  # Should not be triggered due to collision checking
+        # Sanity check if we are below the ground. Should not be triggered due to collision checking
+        below_ground = jnp.any(z_coords < -0.1, axis=1)
         terminated = jnp.logical_or(below_ground, contact)  # no termination condition
         return jnp.where(dones, False, terminated)
 
     @staticmethod
     @jax.jit
     def _truncated(
-        dones: jax.Array, steps: jax.Array, max_episode_steps: jax.Array, n_substeps: jax.Array
-    ) -> jnp.ndarray:
+        dones: Array, steps: Array, max_episode_steps: Array, n_substeps: Array
+    ) -> Array:
         truncated = steps / n_substeps >= max_episode_steps
         return jnp.where(dones, False, truncated)
 
     def render(self):
         self.sim.render()
 
-    def _get_obs(self) -> Dict[str, jnp.ndarray]:
+    def _get_obs(self) -> dict[str, Array]:
         obs = {
-            state: self._maybe_to_numpy(
+            state: maybe_to_numpy(
                 getattr(self.sim.states, state)[..., 2]
                 if state == "pos"
-                else getattr(self.sim.states, state)
+                else getattr(self.sim.states, state),
+                self.return_datatype == "numpy",
             )
             for state in self.states_to_include_in_obs
         }
         return obs
 
-    def _maybe_to_numpy(self, data: Array) -> np.ndarray:
-        if self.return_datatype == "numpy" and not isinstance(data, np.ndarray):
-            return jax.device_get(data)
-        return data
-
 
 class CrazyflowEnvReachGoal(CrazyflowBaseEnv):
     """JAX Gymnasium environment for Crazyflie simulation."""
@@ -284,7 +285,7 @@ def reward(self) -> Array:
 
     @staticmethod
     @jax.jit
-    def _reward(terminated: jax.Array, states: SimState, goal: jax.Array) -> jnp.ndarray:
+    def _reward(terminated: Array, states: SimState, goal: Array) -> Array:
         norm_distance = jnp.linalg.norm(states.pos - goal, axis=2)
         reward = jnp.exp(-2.0 * norm_distance)
         return jnp.where(terminated, -1.0, reward)
@@ -302,7 +303,7 @@ def reset(self, mask: Array) -> None:
         )
         self.goal = self.goal.at[mask].set(new_goals[mask])
 
-    def _get_obs(self) -> Dict[str, jnp.ndarray]:
+    def _get_obs(self) -> dict[str, Array]:
         obs = super()._get_obs()
         obs["difference_to_goal"] = [self.goal - self.sim.states.pos]
         return obs
@@ -329,7 +330,7 @@ def reward(self) -> Array:
 
     @staticmethod
     @jax.jit
-    def _reward(terminated: jax.Array, states: SimState, target_vel: jax.Array) -> jnp.ndarray:
+    def _reward(terminated: Array, states: SimState, target_vel: Array) -> Array:
         norm_distance = jnp.linalg.norm(states.vel - target_vel, axis=2)
         reward = jnp.exp(-norm_distance)
         return jnp.where(terminated, -1.0, reward)
@@ -347,7 +348,7 @@ def reset(self, mask: Array) -> None:
         )
         self.target_vel = self.target_vel.at[mask].set(new_target_vel[mask])
 
-    def _get_obs(self) -> Dict[str, jnp.ndarray]:
+    def _get_obs(self) -> dict[str, Array]:
         obs = super()._get_obs()
         obs["difference_to_target_vel"] = [self.target_vel - self.sim.states.vel]
         return obs