def discount_spec(self) -> t.Tuple[specs.BoundedArray, specs.BoundedArray]:
return (specs.BoundedArray(shape=(),
dtype=float,
minimum=0.,
maximum=1.,
name='P-discount'),
specs.BoundedArray(shape=(),
dtype=float,
minimum=0.,
maximum=1.,
name='D-discount'))
def __init__(self, env):
self._env = env
self._obs_spec = OrderedDict()
wrapped_obs_spec = env.observation_spec().copy()
dim = 0
for key in wrapped_obs_spec.keys():
if key != MANIP_PIXELS_KEY:
spec = wrapped_obs_spec[key]
assert spec.dtype == np.float64
assert type(spec) == specs.Array
dim += np.prod(spec.shape)
self._obs_spec['features'] = specs.Array(shape=(dim, ),
dtype=np.float32,
name='features')
if MANIP_PIXELS_KEY in wrapped_obs_spec:
spec = wrapped_obs_spec[MANIP_PIXELS_KEY]
self._obs_spec['pixels'] = specs.BoundedArray(shape=spec.shape[1:],
dtype=spec.dtype,
minimum=spec.minimum,
maximum=spec.maximum,
name='pixels')
self._obs_spec['state'] = specs.Array(
shape=self._env.physics.get_state().shape,
dtype=np.float32,
name='state')
def __init__(self, scaling_factor=1., action_layers='agent',
constrained_lr=False, control_velocity=False, momentum=0.):
"""Constructor.
Args:
scaling_factor: Scalar. Scaling factor multiplied to the action.
agent_layer: String or iterable of strings. Elements (or itself if
string) must be keys in the environment state. All sprites in
these layers will be acted upon by this action space.
control_velocity: Bool. Whether to control velocity (True) or force
(False).
constrained_lr: Bool. If True, joystick is contrained to actions
parallel to the x-axis, by zeroing out the y-axis (component 1)
of the action.
momentum: Float in [0, 1]. Discount factor for previous action. This
should be zero if control_velocity is False, because imparting
forces automatically gives momentum to the agent(s) being
controlled. If control_velocity is True, setting this greater
than zero gives the controlled agent(s) momentum. However, the
velocity is clipped at scaling_factor, so the agent only retains
momentum when stopping or changing direction and does not
accelerate.
"""
self._scaling_factor = scaling_factor
if not isinstance(action_layers, (list, tuple)):
action_layers = (action_layers,)
self._action_layers = action_layers
self._constrained_lr = constrained_lr
self._control_velocity = control_velocity
self._momentum = momentum
self._action_spec = specs.BoundedArray(
shape=(2,), dtype=np.float32, minimum=-1, maximum=1)
def action_spec(self, physics):
"""Returns a `BoundedArray` matching the `physics` actuators."""
return specs.BoundedArray(shape=(3, ),
dtype=np.float,
minimum=[-1.0, -1.0, -1.0],
maximum=[1.0, 1.0, 1.0])
def action_spec(self):
"""Returns the action spec.
"""
return specs.BoundedArray(
shape=(1,), dtype=int,
minimum=[self._actions[0]], maximum=[self._actions[-1]],
name="buy_sell_action"
)
def action_spec(self, physics):
spec = physics.action_spec()
if self.relative_step:
spec = specs.BoundedArray(shape=(self.DOF,), dtype=np.float,
minimum=np.ones(self.DOF)*-self.relative_rad_max,
maximum=np.ones(self.DOF)*self.relative_rad_max)
return spec
else:
# TODO this will only work if we are using Mujoco - add to robot
spec = physics.action_spec()
if spec.shape[0] == self.DOF:
return spec
# sometimes we only want to control a few joints
elif spec.shape[0] > self.DOF:
return specs.BoundedArray(shape=(self.DOF,), dtype=np.float,
minimum=spec.minimum[:self.DOF],
maximum=spec.maximum[:self.DOF])
else:
raise NotImplemented
