👁️ Observation Space

The observation space defines the information that an agent perceives from the environment.

Building Observations

Basic Observation (Joint State)

def _get_obs(self):
    qpos = self.data.qpos.copy()
    qvel = self.data.qvel.copy()
    return np.concatenate([qpos, qvel]).astype(np.float32)

Extended Observation (+ Sensors)

def _get_obs(self):
    qpos = self.data.qpos.copy()
    qvel = self.data.qvel.copy()

    # IMU data
    imu = self.query_sensor_data(["imu_acc", "imu_gyro"])

    # End-effector pose (returns a nested dict: body_name -> {"xpos", "xmat", "xquat"})
    result = self.get_body_xpos_xmat_xquat(["ee_link"])
    ee_pos = result["ee_link"]["xpos"]  # (3,) end-effector position

    return np.concatenate([
        qpos, qvel,
        imu["imu_acc"], imu["imu_gyro"],
        ee_pos
    ]).astype(np.float32)

Dictionary Observation (Multimodal)

def _get_obs(self):
    return {
        "proprio": np.concatenate([
            self.data.qpos.copy(), 
            self.data.qvel.copy()
        ]).astype(np.float32),
        "vision": self.get_camera_image("front_camera"),  # Requires custom camera capture
        "force": self.query_sensor_data(["ft_sensor"])["ft_sensor"],
    }

Automatic Observation Space Inference

# Automatically inferred on first reset
obs = self._get_obs()
self.observation_space = self.generate_observation_space(obs)

# For numpy observations → spaces.Box
# For dict observations → spaces.Dict

Observation Normalization

class MyEnv(OrcaGymEulerEnv):
    def __init__(self, ...):
        super().__init__(...)

        # Define normalization parameters
        self._obs_mean = np.array([...])  # Task-dependent
        self._obs_std = np.array([...])

    def _get_obs(self):
        raw_obs = np.concatenate([
            self.data.qpos.copy(),
            self.data.qvel.copy()
        ])
        return ((raw_obs - self._obs_mean) / self._obs_std).astype(np.float32)

Observation Construction Best Practices

1. Use copy() to avoid subsequent updates overwriting data
2. Keep float32 for PyTorch/TensorFlow compatibility
3. Fixed observation dimensions — shape must not change during inference
4. Consider historical information — may need to stack multiple frames of observations