reference/robobench/robots/franka.py

"""FrankaRobot — a Franka Emika Panda arm + parallel gripper, for table-top manipulation.

A fixed-base 7-DOF arm with a 2-finger gripper. Five control modes (the gripper is always direct
position targets; switching the mode swaps only the arm controller):

  - "osc"       -> arm by operational-space control (`OperationalSpaceController`, inertia-shaped
                   torque); action = 6 EE pose deltas + 2 gripper = 8. Default — the inertia decoupling
                   keeps the low-inertia wrist smooth.
  - "impedance" -> arm by Jacobian-transpose task-space impedance (`TaskSpaceImpedanceController`);
                   same 8-D action. The form Isaac's Factory tasks use; on this arm it can shake the
                   wrist (no inertia decoupling), so it's not the default.
  - "diff_ik"   -> arm by differential IK (`DiffIKController`, batched DLS step -> joint position
                   targets tracked by the arm PD); same 8-D action (same delta scales as the torque
                   modes). The Isaac Lab task-space example recipe (`FRANKA_PANDA_HIGH_PD_CFG`, arm
                   PD 400/80 — the cfg defaults here).
  - "pink_ik"   -> arm by Pink multi-task QP IK (`PinkIKController`, per-env CPU solve over the
                   vendored kinematics URDF) -> joint position targets tracked by the same arm PD;
                   action = ABSOLUTE hand pose `[pos3, quat4 wxyz]` (env frame) + 2 gripper = 9.
                   The GR00T-school teleop/retarget solver. NOTE: any script building a pink env
                   must `import pinocchio` BEFORE AppLauncher (see `controllers/pink_ik.py`).
  - "joint"     -> arm by direct joint position targets (`JointController`); action = 7 arm + 2 gripper.

The two torque modes load the arm actuators in TORQUE mode (zero stiffness/damping) so the
controller's torques drive them; "diff_ik"/"pink_ik"/"joint" keep the arm position PD
(`arm_stiffness`/`arm_damping`). So `action_dim` and the arm actuator setup follow `control_mode`.
Heavy imports are deferred so registration stays app-free.
"""

from __future__ import annotations

import math
from dataclasses import dataclass
from pathlib import Path
from robobench.core.assets import asset_path
from typing import TYPE_CHECKING, Any, ClassVar

import torch

from robobench.controllers import (
    CompositeController,
    DiffIKController,
    DiffIKControllerCfg,
    FrameTaskCfg,
    JointController,
    JointControllerCfg,
    OperationalSpaceController,
    PinkIKController,
    PinkIKControllerCfg,
    TaskSpaceControllerCfg,
    TaskSpaceImpedanceController,
)
from robobench.core import ROBOTS, BaseRobot, BaseRobotCfg

if TYPE_CHECKING:
    from isaaclab.assets import Articulation

    from robobench.core import BaseEnv


@dataclass
class FrankaRobotCfg(BaseRobotCfg):
    """Config for `FrankaRobot`. `control_mode` is inherited from `BaseRobotCfg` ("" -> first declared).
    `base_pos`/`base_rot` are placement dials; `fixed_base` + the asset are structural. The arm PD gains
    are used only in "joint" mode (OSC zeroes them for torque control); the gripper PD is always on."""

    fixed_base: bool = True  # weld the base to the world (a table-mounted arm)
    base_pos: tuple[float, float, float] = (0.0, 0.0, 0.0)  # base at the table level
    base_rot: tuple[float, float, float, float] = (1.0, 0.0, 0.0, 0.0)  # wxyz; faces +x
    # Arm position-PD gains — used in the position modes ("joint" / "diff_ik"; the torque modes zero
    # them). 400/80 = Isaac's FRANKA_PANDA_HIGH_PD_CFG, "specifically used for IK tracking" — the
    # stock 80/4 preset lags task-space targets badly.
    arm_stiffness: float = 400.0
    arm_damping: float = 80.0
    # Arm actuator effort cap [N*m]; None -> keep the preset's real-Panda limits (87/12). Raise for
    # scripted joint-position tracking that must not crawl at the real limits (kinematic-demo ports).
    arm_effort_limit: float | None = None
    # Body-level gravity compensation fraction (Newton/MuJoCo backend only; 1.0 = weightless arm).
    # The preset's PhysX `disable_gravity` flag is IGNORED by the Newton pipeline — MuJoCo needs
    # `gravcomp`, else a kp=400 servo sags ~tau_g/kp (~0.1 rad on the shoulder when extended).
    # None -> leave the preset untouched (required under the PhysX/2.x venv).
    gravity_compensation: float | None = None
    # PhysX-side gravity compensation: disable gravity on the robot's links (the real Panda
    # gravity-compensates internally; the task-space torque law here has NO gravity term, so
    # with gravity on, extended poses spend the task gains fighting the arm's own weight —
    # MEASURED: commanded reorientations at low work poses stall 56-76 deg from target while
    # the same rotations track to 1-7 deg from the light tuck pose).
    disable_arm_gravity: bool = False
    # Gripper PD gains (always position-controlled; holds / grasps the part).
    gripper_stiffness: float = 2000.0
    gripper_damping: float = 100.0
    # Gripper actuator effort cap [N]; None -> keep the preset's default. Raise for pinch grips
    # that must not saturate (e.g. 500.0 for cloth, the isaaclab soft-lift tasks' value).
    gripper_effort_limit: float | None = None
    # Home posture of the 7 arm joints: the arm resets here. A forward-facing ready pose; retune per
    # task (e.g. to start the gripper near the work).
    default_dof_pos: tuple[float, ...] = (0.0015, -0.197, -0.0014, -1.976, -0.00028, 1.78, 0.786)
    # Posture the task-space nullspace pulls toward; () -> use default_dof_pos. A non-singular elbow
    # config that actively resolves the arm's redundancy (keeps the wrist from drifting); kept separate
    # from the home pose on purpose.
    nullspace_dof_pos: tuple[float, ...] = (-1.3003, -0.4015, 1.1791, -2.1493, 0.4001, 1.9425, 0.4754)
    # Nullspace posture stiffness; None -> the controller's default (10.0). MEASURED: at 10.0
    # the posture spring overpowers the task torque for world-z reorientations (stalls 76 deg
    # from target with every joint far from its limits) — soften it when a task needs large
    # wrist reorientation. Damping follows critically (2*sqrt(kp)).
    kp_null: float | None = None
    # Task-space stiffness override [xyz, rpy]; None -> the controller's default
    # (100,100,100, 30,30,30). MEASURED: at rot-gain 30 a wrist welded to a ~2.5 kg
    # payload rolls ~16 deg toward a commanded 100 deg reorientation and equilibrates
    # (free-space: 1 deg residual) — the rotation channel needs more authority under
    # payload coupling.
    task_prop_gains: tuple[float, ...] | None = None
    franka_usd: str = ""  # "" -> the vendored robots/assets/franka/panda_instanceable.usd
    franka_urdf: str = ""  # "" -> the vendored kinematics URDF (used by the pink_ik control mode)

    def __post_init__(self) -> None:
        assets = asset_path(Path(__file__).resolve().parent / "assets") / "franka"
        self.franka_usd = self.franka_usd or str(assets / "panda_instanceable.usd")
        self.franka_urdf = self.franka_urdf or str(assets / "panda_kinematics.urdf")


@ROBOTS.register("franka")
class FrankaRobot(BaseRobot):
    """Franka Panda arm + gripper. `apply_action` delegates to the controller for `control_mode`; the
    2 gripper fingers are always direct position targets, the 7 arm joints by the mode's arm controller
    (torque-mode OSC, or position-mode JointController)."""

    control_modes: tuple[str, ...] = ("osc", "impedance", "diff_ik", "pink_ik", "joint")  # osc default: smooth on this arm
    cfg: FrankaRobotCfg

    ARM_JOINTS: tuple[str, ...] = ("panda_joint[1-7]",)
    GRIPPER_JOINTS: tuple[str, ...] = ("panda_finger_joint.*",)
    EE_BODY: str = "panda_hand"  # the OSC control frame (a real body with a Jacobian)

    #: Eye-in-hand view at the standard RealSense D435 bracket pose (ManiSkill panda_wristcam /
    #: DROID rig): beside the hand, looking parallel to the +z approach axis, ~90 deg FOV.
    CAMERAS: ClassVar[dict[str, dict]] = {
        "wrist": {"link": "panda_hand", "eye": (0.0465, -0.02, 0.036), "target": (0.0465, -0.02, 0.3),
                  "focal": 10.5},
    }

    # Action/target rate (s). Torque modes (osc/impedance): ~15 Hz target (Isaac `Factory-NutThread`,
    # decim 8 @ 120 Hz), latched while the torque law recomputes every physics step (see
    # `_TaskSpaceController.apply`). Joint mode: ~50 Hz, held by the actuator PD.
    TORQUE_CONTROL_DT: float = 1.0 / 15.0
    JOINT_CONTROL_DT: float = 0.02

    def __init__(self, cfg: FrankaRobotCfg | None = None) -> None:
        super().__init__(cfg or FrankaRobotCfg())

    # ----- assets -------------------------------------------------------------------------------
    def assets(self) -> dict[str, Any]:
        """The Franka articulation (`FRANKA_PANDA_HIGH_PD_CFG`), base fixed per `cfg.fixed_base`, spawned
        at the configured pose. In the torque modes ("impedance"/"osc") the arm actuators are set to
        torque mode (zero stiffness/damping); in "diff_ik"/"joint" mode they keep position PD. The gripper is always PD."""
        from isaaclab_assets.robots.franka import FRANKA_PANDA_HIGH_PD_CFG

        c = self.cfg
        robot = FRANKA_PANDA_HIGH_PD_CFG.copy()  # type: ignore[attr-defined]
        robot.prim_path = f"{{ENV_REGEX_NS}}/{self.prim_name}"  # cfg.name-namespaced (default "Robot")
        robot.spawn.usd_path = c.franka_usd  # vendored local panda (resolved in cfg.__post_init__)
        robot.spawn.articulation_props.fix_root_link = c.fixed_base
        robot.init_state.pos = c.base_pos
        robot.init_state.rot = c.base_rot
        # Home the 7 arm joints to cfg.default_dof_pos (reset() reads this via default_joint_pos);
        # keep the gripper-finger defaults.
        robot.init_state.joint_pos = {
            **robot.init_state.joint_pos,
            **{f"panda_joint{i + 1}": float(q) for i, q in enumerate(c.default_dof_pos)},
        }
        torque_mode = self.control_mode in ("impedance", "osc")
        for arm_act in ("panda_shoulder", "panda_forearm"):
            robot.actuators[arm_act].stiffness = 0.0 if torque_mode else c.arm_stiffness
            robot.actuators[arm_act].damping = 0.0 if torque_mode else c.arm_damping
            if c.arm_effort_limit is not None:
                robot.actuators[arm_act].effort_limit_sim = c.arm_effort_limit
        robot.actuators["panda_hand"].stiffness = c.gripper_stiffness
        robot.actuators["panda_hand"].damping = c.gripper_damping
        if c.gripper_effort_limit is not None:
            robot.actuators["panda_hand"].effort_limit_sim = c.gripper_effort_limit
        if c.gravity_compensation is not None:
            # Newton-backend gravity compensation: swap in the MuJoCo rigid-body schema carrying
            # `gravcomp` (imported lazily so the PhysX/2.x venv never touches isaaclab_newton).
            from isaaclab_newton.sim.schemas.schemas_cfg import MujocoRigidBodyPropertiesCfg

            robot.spawn.rigid_props = MujocoRigidBodyPropertiesCfg(gravcomp=c.gravity_compensation)
        if c.disable_arm_gravity:
            robot.spawn.rigid_props.disable_gravity = True
        return {self.name: robot}

    # ----- lifecycle (hooks; the base orchestrates bind -> on_bind -> build_controller) ---------
    def on_bind(self, env: BaseEnv) -> None:
        self.articulation: Articulation = env.iscene[self.name]

    def build_controller(self) -> CompositeController:
        """`composite([<arm controller>, joint(gripper)])` for the active mode. The gripper is always a
        position JointController (2 fingers); the arm controller is OSC (torque) or JointController
        (position)."""
        torque_mode = self.control_mode in ("impedance", "osc")
        ctrl_dt = self.TORQUE_CONTROL_DT if torque_mode else self.JOINT_CONTROL_DT
        gripper = JointController(JointControllerCfg(self.GRIPPER_JOINTS, dt=ctrl_dt), command_type="position")
        if torque_mode:
            ts_cfg = TaskSpaceControllerCfg(
                dt=ctrl_dt,  # target rate; the torque law itself recomputes every physics step (see apply)
                ee_body=self.EE_BODY,
                arm_joint_names=self.ARM_JOINTS,
                nullspace_dof_pos=self.cfg.nullspace_dof_pos or self.cfg.default_dof_pos,  # () -> home pose
                ema_factor=0.2,  # smooth the action stream
            )
            if self.cfg.kp_null is not None:
                ts_cfg.kp_null = self.cfg.kp_null
                ts_cfg.kd_null = 2.0 * math.sqrt(self.cfg.kp_null)
            if self.cfg.task_prop_gains is not None:
                ts_cfg.task_prop_gains = tuple(self.cfg.task_prop_gains)
                ts_cfg.task_deriv_gains = ()  # keep critical damping (2*sqrt(kp))
            cls = TaskSpaceImpedanceController if self.control_mode == "impedance" else OperationalSpaceController
            arm: Any = cls(ts_cfg)
        elif self.control_mode == "diff_ik":
            # Same 6-D delta action as the torque modes; the DLS step emits joint position targets
            # the arm PD (cfg.arm_stiffness/arm_damping, HIGH_PD 400/80) tracks.
            arm = DiffIKController(DiffIKControllerCfg(dt=ctrl_dt, ee_body=self.EE_BODY, arm_joint_names=self.ARM_JOINTS))
        elif self.control_mode == "pink_ik":
            # One frame task on the hand over the 7-joint chain (kinematics-only URDF; USD names ==
            # URDF names on the panda). Task costs = the G1's proven set (position 8/m, orientation
            # 2/rad, gain 0.5, lm 10 — the FrameTaskCfg defaults); the 1-DOF elbow redundancy is
            # anchored by the nullspace posture pull toward home. Emits joint position targets the
            # same 400/80 arm PD tracks.
            arm = PinkIKController(
                PinkIKControllerCfg(
                    dt=ctrl_dt,
                    urdf_path=self.cfg.franka_urdf,
                    base_link="panda_link0",
                    base_link_frame="panda_link0",
                    frames=(FrameTaskCfg(self.EE_BODY),),
                    joint_names=self.ARM_JOINTS,
                    nullspace_joints=tuple(f"panda_joint{i}" for i in range(1, 8)),
                )
            )
        elif self.control_mode == "joint":
            arm = JointController(JointControllerCfg(self.ARM_JOINTS, dt=ctrl_dt), command_type="position")
        else:
            raise ValueError(f"unknown Franka control_mode {self.control_mode!r}; known: {self.control_modes}")
        return CompositeController([arm, gripper])

    def reset(self, env_ids: torch.Tensor) -> None:
        """Home pose: default joint state; hold all joint position targets at default (the gripper PD
        holds, the arm targets are inert in torque mode). Root written to the fixed spawn pose + origin."""
        art = self.articulation
        jp = art.data.default_joint_pos[env_ids].clone()
        jv = art.data.default_joint_vel[env_ids].clone()
        art.write_joint_state_to_sim(jp, jv, env_ids=env_ids)
        art.set_joint_position_target(jp, env_ids=env_ids)
        art.set_joint_effort_target(torch.zeros_like(jp), env_ids=env_ids)
        root = art.data.default_root_state[env_ids].clone()
        root[:, 0:3] += self.env.iscene.env_origins[env_ids]
        art.write_root_state_to_sim(root, env_ids)
        if self.controller is not None:
            self.controller.reset(env_ids)

    # NOTE: apply_action / action_dim are inherited from BaseRobot — they delegate to self.controller
    # (composite of arm + gripper). The OSC arm writes effort, the gripper writes position, each through
    # the sink it captured at bind.

    # get_state / set_state are inherited from BaseRobot (single-articulation sim state + controller).

    # ----- description --------------------------------------------------------------------------
    def describe(self) -> str:
        mode = self.control_mode
        if mode == "impedance":
            arm = "7 arm joints by task-space impedance (joint torque); the action is 6 end-effector pose deltas"
        elif mode == "osc":
            arm = "7 arm joints by operational-space control (joint torque); the action is 6 end-effector pose deltas"
        elif mode == "diff_ik":
            arm = "7 arm joints by differential IK (joint position targets); the action is 6 end-effector pose deltas"
        elif mode == "pink_ik":
            arm = ("7 arm joints by Pink QP IK (joint position targets); the action is the absolute hand pose "
                   "[pos, quat wxyz] in the env frame")
        else:
            arm = "7 arm joints by direct position targets"
        return (
            f"A Franka Emika Panda arm with a parallel-jaw gripper, fixed to the table. Control mode "
            f"'{mode}': {arm}, plus 2 gripper fingers by direct position target. Action dim {self.action_dim}."
        )