simulation/scripts/play_coordex_v2d.py

#!/usr/bin/env python3
"""Roll out the released CoordEx WalkGrab policy on the v2d reconstructed object.

This is not SONIC and not ``V2D-G1-SonicManip``. CoordEx's actor consumes
Wuji proprio history, VAE prior means, cube/table poses, and fingertip forces,
and emits 28 residual latents. Those tensors do not exist on the 4-D wrist+grip
task, so the weights cannot be loaded there.

The checkpoint was trained with rsl-rl 2.x, whose ``OnPolicyRunner`` cannot read
this repo's rsl-rl 5.x config schema. The actor is therefore built and loaded
directly, and the env is stepped through its Gym API instead.

  cd simulation
  ./jobs/coordex_v2d_play.sh --headless --video
"""

from __future__ import annotations

import argparse
import os
import pathlib
import sys
from math import prod

_PROJECT = pathlib.Path(__file__).resolve().parents[2]
_COORDEX = _PROJECT / "third-party" / "coordex"
_COORDEX_PKG = _COORDEX / "source" / "coordex"
for p in (_COORDEX_PKG, _PROJECT / "simulation" / "source"):
    s = str(p)
    if s not in sys.path:
        sys.path.insert(0, s)

from isaaclab.app import AppLauncher

parser = argparse.ArgumentParser(
    description="CoordEx WalkGrab policy on the v2d coffee-can (or exported URDF)."
)
parser.add_argument("--task", default="V2D-CoorDex-WalkGrab-Play-v0")
parser.add_argument(
    "--checkpoint",
    default=str(_COORDEX / "ckpts/locomanip/walkgrab_8k.pt"),
)
parser.add_argument(
    "--body-prior-checkpoint",
    default=str(_COORDEX / "ckpts/body_prior/walkgrab_10k.pt"),
)
parser.add_argument(
    "--hand-prior-checkpoint",
    default=str(_COORDEX / "ckpts/hand_prior/kinematic_wrist_16k.pt"),
)
parser.add_argument(
    "--clip",
    default="20200709_141754_836212060125",
    help="reconstruction/runs/<clip> whose exported object.urdf goes on the table.",
)
parser.add_argument("--num_envs", type=int, default=1)
parser.add_argument("--max_steps", type=int, default=0, help="0 = one full episode.")
parser.add_argument("--video", action="store_true")
parser.add_argument(
    "--video_length",
    type=int,
    default=0,
    help="Recorded steps; 0 = whole run. Used when --video is set.",
)
parser.add_argument(
    "--norm-eps",
    type=float,
    default=1.0e-2,
    help="rsl-rl 2.x EmpiricalNormalization eps; the checkpoint was written with the default.",
)
AppLauncher.add_app_launcher_args(parser)
args_cli, hydra_args = parser.parse_known_args()
sys.argv = [sys.argv[0], *hydra_args]
# Read by g1_coordex_v2d_env_cfg.object_urdf() when the cfg is built.
os.environ["V2D_COORDEX_CLIP"] = args_cli.clip
if args_cli.video:
    args_cli.enable_cameras = True

app_launcher = AppLauncher(args_cli)
simulation_app = app_launcher.app

import gymnasium as gym  # noqa: E402
import torch  # noqa: E402
from coordex.policies import ActorCriticCoordResidual  # noqa: E402
from coordex.tasks.locomanip.constants import RIGHT_HAND_TIP_NAMES  # noqa: E402
from isaaclab.utils.dict import print_dict  # noqa: E402
from isaaclab_tasks.utils.hydra import hydra_task_config  # noqa: E402

import v2d_sim.tasks.coordex_v2d  # noqa: E402, F401
from v2d_sim.tasks.coordex_v2d.proxy_compat import apply_data_tensor_compat  # noqa: E402


def _resolve_path(path: str | os.PathLike[str]) -> str:
    candidate = pathlib.Path(path).expanduser()
    if not candidate.is_absolute():
        candidate = _COORDEX / candidate
    candidate = candidate.resolve()
    if not candidate.is_file():
        raise FileNotFoundError(f"Checkpoint not found: {candidate}")
    return str(candidate)


def _t(x):
    """Isaac Lab 6 hands back warp ProxyArray from ``.data``; CoordEx wants Tensors."""
    return x.torch if hasattr(x, "torch") else x


def _term_dim_to_int(dim) -> int:
    if isinstance(dim, int):
        return int(dim)
    if isinstance(dim, tuple):
        return int(prod(dim))
    return int(prod(int(value) for value in dim))


def _actor_obs_metadata(env) -> tuple[tuple[str, ...], list[int]]:
    obs_mgr = getattr(env.unwrapped, "observation_manager", None)
    if obs_mgr is None:
        raise RuntimeError("CoordResidual policy requires a ManagerBased observation manager.")
    names = tuple(obs_mgr._group_obs_term_names["policy"])
    dims = [_term_dim_to_int(dim) for dim in obs_mgr._group_obs_term_dim["policy"]]
    return names, dims


def _group_dim(env, group: str) -> int:
    return int(prod(env.unwrapped.observation_manager.group_obs_dim[group]))


def _build_actor(env, agent_cfg, checkpoint: str, device: str):
    """Instantiate ActorCriticCoordResidual and load the rsl-rl 2.x checkpoint."""
    policy_cfg = agent_cfg.policy
    names, dims = _actor_obs_metadata(env)
    num_actor_obs = _group_dim(env, "policy")
    num_critic_obs = _group_dim(env, "critic")
    num_actions = int(env.unwrapped.action_manager.total_action_dim)

    model = ActorCriticCoordResidual(
        num_actor_obs=num_actor_obs,
        num_critic_obs=num_critic_obs,
        num_actions=num_actions,
        critic_hidden_dims=policy_cfg.critic_hidden_dims,
        activation=policy_cfg.activation,
        init_noise_std=policy_cfg.init_noise_std,
        noise_std_type=policy_cfg.noise_std_type,
        coord_trunk_hidden_dims=policy_cfg.coord_trunk_hidden_dims,
        body_head_hidden_dims=policy_cfg.body_head_hidden_dims,
        hand_head_hidden_dims=policy_cfg.hand_head_hidden_dims,
        body_residual_scale=policy_cfg.body_residual_scale,
        hand_residual_scale=policy_cfg.hand_residual_scale,
        fixed_log_std=policy_cfg.fixed_log_std,
        actor_obs_term_names=names,
        actor_obs_term_dims=dims,
    ).to(device)

    payload = torch.load(checkpoint, map_location=device, weights_only=False)
    ckpt_in = payload["model_state_dict"]["actor.coord_trunk.0.weight"].shape[1]
    live_in = model.actor.coord_trunk[0].weight.shape[1]
    if ckpt_in != live_in:
        raise RuntimeError(
            f"coord trunk input mismatch: checkpoint {ckpt_in} vs env {live_in}.\n"
            f"policy obs dim={num_actor_obs}, terms={list(zip(names, dims))}\n"
            "The env's observation layout drifted from the one WalkGrab was trained on."
        )
    model.load_state_dict(payload["model_state_dict"], strict=True)
    model.eval()

    norm = payload.get("obs_norm_state_dict")
    if norm is None:
        mean = torch.zeros(num_actor_obs, device=device)
        std = torch.ones(num_actor_obs, device=device)
    else:
        mean = norm["_mean"].to(device).reshape(-1)
        std = norm["_std"].to(device).reshape(-1)
    if mean.numel() != num_actor_obs:
        raise RuntimeError(
            f"obs normalizer is {mean.numel()}-D but the policy group is {num_actor_obs}-D."
        )

    eps = float(args_cli.norm_eps)

    def policy(obs: torch.Tensor) -> torch.Tensor:
        return model.act_inference((_t(obs) - mean) / (std + eps))

    print(
        f"[coordex-v2d] actor obs={num_actor_obs} critic obs={num_critic_obs} "
        f"actions={num_actions} (checkpoint iter={payload.get('iter')})",
        flush=True,
    )
    return policy


def _iter_prims(prim):
    from pxr import Usd

    yield prim
    # Colliders often sit inside instanced references, which GetChildren() skips.
    for child in prim.GetFilteredChildren(Usd.TraverseInstanceProxies(Usd.PrimAllPrimsPredicate)):
        yield from _iter_prims(child)


def _dump_object_prims(root: str = "/World/envs/env_0/Bottle") -> None:
    """Print the physics schemas under the Bottle prim.

    The fingertip sensors filter on exactly this path. If the rigid body and collider
    actually live on a child prim (the URDF importer nests link prims), or if the
    contact report API never got applied, PhysX cannot build the filter and
    ``force_matrix_w`` stays zero. Both cases are visible here.
    """
    import omni.usd
    from pxr import PhysxSchema, UsdPhysics

    stage = omni.usd.get_context().get_stage()
    prim = stage.GetPrimAtPath(root)
    if not prim or not prim.IsValid():
        print(f"[coordex-v2d] prim dump: {root} not found", flush=True)
        return
    print(f"[coordex-v2d] prim tree under {root}:", flush=True)
    for p in _iter_prims(prim):
        tags = []
        if p.HasAPI(UsdPhysics.RigidBodyAPI):
            tags.append("RigidBody")
        if p.HasAPI(UsdPhysics.CollisionAPI):
            tags.append("Collision")
        if p.HasAPI(UsdPhysics.MeshCollisionAPI):
            approx = UsdPhysics.MeshCollisionAPI(p).GetApproximationAttr().Get()
            tags.append(f"MeshCollision({approx})")
        if p.HasAPI(PhysxSchema.PhysxContactReportAPI):
            tags.append("ContactReport")
        print(
            f"    {p.GetPath()}  type={p.GetTypeName() or '-'}  {' '.join(tags) or '-'}",
            flush=True,
        )


_TIP_SENSORS = (
    "contact_force_thumb",
    "contact_force_index",
    "contact_force_middle",
    "contact_force_ring",
    "contact_force_pinky",
)


class _ContactProbe:
    """Track whether the fingertip->Bottle contact filter actually reports anything.

    WalkGrab's ``fingertip_cube_forces`` term reads ``force_matrix_w`` from five
    sensors filtered to the Bottle prim. If PhysX cannot build that filter (it warns
    ``GPU contact filter for collider ... is not supported``) the field stays zero and
    the term silently feeds the policy 15 zeros, i.e. "never touching anything".
    Without this probe a dead grasp input is indistinguishable from a bad policy.
    """

    def __init__(self, env) -> None:
        sensors = getattr(env.unwrapped.scene, "sensors", None) or {}
        self.sensors = {n: sensors[n] for n in _TIP_SENSORS if n in sensors}
        self.missing = [n for n in _TIP_SENSORS if n not in sensors]
        self.field: str | None = None
        self.peak = 0.0
        self.peak_unfiltered = 0.0
        for sensor in self.sensors.values():
            for candidate in ("force_matrix_w", "force_matrix_w_history", "net_forces_w"):
                if getattr(sensor.data, candidate, None) is not None:
                    self.field = candidate
                    break
            break

    def describe(self) -> str:
        if not self.sensors:
            return "[coordex-v2d] WARNING no fingertip contact sensors on the scene"
        msg = f"[coordex-v2d] fingertip sensors={len(self.sensors)} field={self.field}"
        if self.missing:
            msg += f" missing={self.missing}"
        return msg

    def update(self) -> float:
        step_peak = 0.0
        for sensor in self.sensors.values():
            value = getattr(sensor.data, self.field, None) if self.field else None
            if value is not None:
                step_peak = max(step_peak, float(_t(value).abs().max()))
            # Unfiltered force distinguishes "the filter is broken" from "the fingers
            # genuinely touched nothing": net_forces_w counts contact with anything.
            net = getattr(sensor.data, "net_forces_w", None)
            if net is not None:
                self.peak_unfiltered = max(self.peak_unfiltered, float(_t(net).abs().max()))
        self.peak = max(self.peak, step_peak)
        return step_peak


@hydra_task_config(args_cli.task, "rsl_rl_cfg_entry_point")
def main(env_cfg, agent_cfg):
    checkpoint = _resolve_path(args_cli.checkpoint)
    body_prior = _resolve_path(args_cli.body_prior_checkpoint)
    hand_prior = _resolve_path(args_cli.hand_prior_checkpoint)

    if args_cli.num_envs is not None:
        env_cfg.scene.num_envs = args_cli.num_envs
    if getattr(args_cli, "device", None) is not None:
        env_cfg.sim.device = args_cli.device
        agent_cfg.device = args_cli.device

    joint_action_cfg = getattr(getattr(env_cfg, "actions", None), "joint_pos", None)
    if joint_action_cfg is None:
        raise RuntimeError("Locomanip env is missing actions.joint_pos.")
    joint_action_cfg.body_prior_checkpoint = body_prior
    joint_action_cfg.hand_prior_checkpoint = hand_prior

    print(f"[coordex-v2d] task={args_cli.task}", flush=True)
    print(f"[coordex-v2d] policy {checkpoint}", flush=True)
    print(f"[coordex-v2d] body prior {body_prior}", flush=True)
    print(f"[coordex-v2d] hand prior {hand_prior}", flush=True)
    print(
        "[coordex-v2d] object is the v2d URDF at WalkGrab bottle xy=(1.5,-0.35); "
        "robot is G1-Wuji, not Inspire+SONIC",
        flush=True,
    )

    env = gym.make(
        args_cli.task,
        cfg=env_cfg,
        render_mode="rgb_array" if args_cli.video else None,
    )
    apply_data_tensor_compat(env)

    # Default to exactly one episode rather than a round number, so a run is never
    # cut off mid-attempt. WalkGrab is 15 s at 60 Hz, i.e. 900 steps, not 500.
    episode_steps = int(env.unwrapped.max_episode_length)
    n = args_cli.max_steps or episode_steps
    video_length = args_cli.video_length or n
    print(f"[coordex-v2d] episode={episode_steps} steps, running {n}, recording {video_length}", flush=True)

    if args_cli.video:
        tag = "stock" if "Stock" in args_cli.task else args_cli.clip
        out_dir = _PROJECT / "simulation" / "runs" / "coordex_v2d" / tag
        out_dir.mkdir(parents=True, exist_ok=True)
        video_kwargs = {
            "video_folder": str(out_dir),
            "step_trigger": lambda step: step == 0,
            "video_length": video_length,
            "disable_logger": True,
        }
        print("[coordex-v2d] recording video", flush=True)
        print_dict(video_kwargs, nesting=4)
        env = gym.wrappers.RecordVideo(env, **video_kwargs)

    device = env.unwrapped.device
    policy = _build_actor(env, agent_cfg, checkpoint, device)

    obs_dict, _ = env.reset()
    min_root_z = float("inf")
    robot = env.unwrapped.scene["robot"]
    obj = env.unwrapped.scene["bottle"]
    palm_id = robot.find_bodies(["right_palm_link"], preserve_order=True)[0][0]
    tip_ids = robot.find_bodies(list(RIGHT_HAND_TIP_NAMES), preserve_order=True)[0]
    probe = _ContactProbe(env)
    print(probe.describe(), flush=True)
    _dump_object_prims()

    term_mgr = env.unwrapped.termination_manager
    resets: dict[str, int] = {}

    # Baseline after the object has settled: spawning can leave it slightly
    # interpenetrating, and measuring lift from the spawn z inflates it.
    settle_step = 10
    obj_z0 = float(_t(obj.data.root_pos_w)[0, 2])
    max_lift = 0.0
    min_palm_dist = float("inf")
    min_tip_dist = float("inf")
    for step in range(n):
        if not simulation_app.is_running():
            break
        with torch.inference_mode():
            actions = policy(obs_dict["policy"])
            obs_dict, _, terminated, truncated, _ = env.step(actions)
        if bool(terminated[0]) or bool(truncated[0]):
            fired = [t for t in term_mgr.active_terms if bool(term_mgr.get_term(t)[0])]
            for name in fired or ["<unknown>"]:
                resets[name] = resets.get(name, 0) + 1
            print(f"  step {step}: episode reset by {fired or '<unknown>'}", flush=True)
        root_z = float(_t(robot.data.root_pos_w)[0, 2])
        min_root_z = min(min_root_z, root_z)
        obj_pos = _t(obj.data.root_pos_w)[0]
        if step == settle_step:
            obj_z0 = float(obj_pos[2])
        max_lift = max(max_lift, float(obj_pos[2]) - obj_z0)
        bodies = _t(robot.data.body_pos_w)[0]
        palm_dist = float(torch.linalg.norm(bodies[palm_id] - obj_pos))
        min_palm_dist = min(min_palm_dist, palm_dist)
        tip_dist = float(torch.linalg.norm(bodies[tip_ids] - obj_pos, dim=-1).min())
        min_tip_dist = min(min_tip_dist, tip_dist)
        force = probe.update()
        if step % 50 == 0:
            print(
                f"  step {step}/{n}  root_z={root_z:.3f}  object_z={float(obj_pos[2]):.3f}"
                f"  palm_dist={palm_dist:.3f}  tip_force={force:.3f}",
                flush=True,
            )

    print(f"[coordex-v2d] min root_z={min_root_z:.3f}", flush=True)
    print(f"[coordex-v2d] min palm-object distance={min_palm_dist:.3f} m", flush=True)
    print(f"[coordex-v2d] min fingertip-object distance={min_tip_dist:.3f} m", flush=True)
    print(f"[coordex-v2d] peak object lift above rest={max_lift:+.3f} m", flush=True)
    print(
        f"[coordex-v2d] peak fingertip force: filtered={probe.peak:.4f} "
        f"unfiltered={probe.peak_unfiltered:.4f}",
        flush=True,
    )
    if probe.peak == 0.0 and probe.peak_unfiltered > 0.0:
        print(
            "[coordex-v2d] WARNING fingertips registered contact with something but the "
            "Bottle filter reported nothing: the filter path is still wrong.",
            flush=True,
        )
    print(f"[coordex-v2d] episode resets: {resets or 'none'}", flush=True)
    if resets.get("robot_passed_bottle_without_grasp"):
        print(
            "[coordex-v2d] NOTE robot_passed_bottle_without_grasp fires when the robot "
            "walks past the object and is_grasped() is False. is_grasped() requires "
            "fingertip contact, so a dead contact filter forces this termination "
            "regardless of how well the policy actually reached.",
            flush=True,
        )
    if probe.peak == 0.0 and min_palm_dist < 0.15:
        print(
            "[coordex-v2d] WARNING hand reached the object but the contact filter never "
            "reported force: fingertip_cube_forces fed the policy 15 zeros, so this run "
            "does not test CoordEx's grasp input fairly.",
            flush=True,
        )
    env.close()


if __name__ == "__main__":
    try:
        main()
    except Exception:
        import traceback

        traceback.print_exc()
        sys.exit(1)
    finally:
        simulation_app.close()