v2d / simulation /scripts /smoke_sonic_manip.py

Upload v2d project (excluding data and .venv) (part 18)

3f4bb1d verified 2 months ago

8.57 kB

#!/usr/bin/env python3
"""Drive V2D-G1-SonicManip with a scripted pick, before any training.
A learned policy failing tells you nothing about whether the environment is
sound. This runs an open-loop reach / close / lift schedule instead, so the
question is narrow: with a sensible action sequence, does the robot stay up,
does SONIC put the hand where it was told, and does the object come off the
table? Anything that fails here would have failed silently during training.
  cd simulation
  ./run_isaaclab.sh --python scripts/smoke_sonic_manip.py --headless
"""

from __future__ import annotations

import argparse
import os
import sys
import traceback
from pathlib import Path

from isaaclab.app import AppLauncher

if hasattr(sys.stdout, "reconfigure"):
    sys.stdout.reconfigure(line_buffering=True)

parser = argparse.ArgumentParser(description="Scripted-pick smoke test for the SONIC manip env.")
parser.add_argument("--task", type=str, default="V2D-G1-SonicManip-Play-v0")
parser.add_argument("--num_envs", type=int, default=4)
parser.add_argument("--steps", type=int, default=150)
parser.add_argument("--video", type=Path, default=None, help="optional MP4 of env 0")
parser.add_argument("--fps", type=float, default=25.0)
AppLauncher.add_app_launcher_args(parser)
args_cli = parser.parse_args()
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 numpy as np  # noqa: E402
import torch  # noqa: E402
from isaaclab.utils.math import quat_apply  # noqa: E402
from isaaclab_tasks.utils import parse_env_cfg  # noqa: E402

import v2d_sim  # noqa: E402, F401
from v2d_sim.tasks.g1_sonic_manip.mdp.observations import t as _t  # noqa: E402

# Observation layout from PolicyObsCfg, in declaration order.
_GRAV, _WRIST, _OBJ, _DELTA, _OBJV, _GRIP = (
    slice(0, 3),
    slice(3, 6),
    slice(6, 9),
    slice(9, 12),
    slice(12, 15),
    slice(15, 16),
)


#: Long enough for a marginally balanced object to visibly roll or topple.
_SETTLE_STEPS = 20


def _object_repose(obj) -> tuple[torch.Tensor, torch.Tensor]:
    """Object position and how far its up axis leans off vertical, in degrees.
    The can's up axis is body +y: the reconstructed mesh is y-up, which is the
    whole reason its spawn rotation matters.
    """
    pos = _t(obj.data.root_pos_w).clone()
    quat = _t(obj.data.root_quat_w)
    body_up = torch.tensor([0.0, 1.0, 0.0], device=quat.device).expand(quat.shape[0], 3)
    up_w = quat_apply(quat, body_up)
    return pos, torch.rad2deg(torch.arccos(up_w[:, 2].clamp(-1.0, 1.0)))


def _rgb(img) -> np.ndarray | None:
    if img is None:
        return None
    if hasattr(img, "detach"):
        img = img.detach().cpu().numpy()
    img = np.asarray(img)
    if img.ndim == 4:
        img = img[0]
    img = img[..., :3]
    if img.dtype != np.uint8:
        scale = 255.0 if float(np.nanmax(img)) <= 1.5 else 1.0
        img = np.clip(img * scale, 0, 255).astype(np.uint8)
    h, w = img.shape[:2]
    return img[: h - (h % 2), : w - (w % 2)]


def main() -> None:
    env_cfg = parse_env_cfg(args_cli.task, device=args_cli.device, num_envs=args_cli.num_envs)
    render = "rgb_array" if args_cli.video else None
    env = gym.make(args_cli.task, cfg=env_cfg, render_mode=render)
    inner = env.unwrapped
    obs_dict, _ = env.reset()
    obs = obs_dict["policy"]

    acfg = env_cfg.actions.wrist
    lo = torch.tensor(
        [acfg.wrist_x[0], acfg.wrist_y[0], acfg.wrist_z[0]], device=inner.device
    )
    hi = torch.tensor(
        [acfg.wrist_x[1], acfg.wrist_y[1], acfg.wrist_z[1]], device=inner.device
    )
    center, half = (lo + hi) / 2.0, (hi - lo) / 2.0

    robot = inner.scene["robot"]
    obj = inner.scene["object"]
    table_top = env_cfg.scene.table.init_state.pos[2] + 0.5 * env_cfg.scene.table.spawn.size[2]
    # Height gained is measured from the resting pose, not the table top: the
    # object origin already sits half an object-height above the table.
    rest_z = env_cfg.scene.object.init_state.pos[2]

    print(f"[smoke] task={args_cli.task}  envs={inner.num_envs}  device={inner.device}")
    print(f"[smoke] obs dim={obs.shape[1]}  action dim={inner.action_manager.total_action_dim}")
    print(f"[smoke] joints={len(robot.joint_names)}  table_top_z={table_top:.3f}")
    print(f"[smoke] object rests at z={rest_z:.3f} ({rest_z - table_top:.3f} above the table)")
    print(f"[smoke] wrist box x{tuple(acfg.wrist_x)} y{tuple(acfg.wrist_y)} z{tuple(acfg.wrist_z)}")
    print(f"[smoke] object start (pelvis frame) = {obs[0, _OBJ].tolist()}")

    # Does the can sit still before anything touches it? A wrong spawn rotation
    # leaves it balanced on a curved side, which the run summary cannot
    # distinguish from a grasp that knocked it over.
    p0, tilt0 = _object_repose(obj)
    hold = torch.zeros(inner.num_envs, 4, device=inner.device)
    hold[:, 2] = 1.0  # wrist parked at the top of its box, clear of the object
    hold[:, 3] = -1.0
    for _ in range(_SETTLE_STEPS):
        env.step(hold)
    p1, tilt1 = _object_repose(obj)
    drift = torch.norm((p1 - p0)[:, :2], dim=1)
    print(
        f"[smoke] untouched for {_SETTLE_STEPS} steps: xy drift {drift.max():.4f} m, "
        f"tilt {tilt0.max():.1f} -> {tilt1.max():.1f} deg from vertical "
        "(firm if drift < ~5 mm and tilt stays put)"
    )
    obs_dict, _ = env.reset()
    obs = obs_dict["policy"]

    n = args_cli.steps
    reach_end, close_end = int(0.40 * n), int(0.55 * n)
    frames: list[np.ndarray] = []
    log: list[tuple[int, float, float, float, float, float]] = []
    min_root_z = float("inf")

    for i in range(n):
        if not simulation_app.is_running():
            break
        object_b = obs[:, _OBJ]
        want = object_b.clone()
        if i < reach_end:
            # Approach from slightly above, so the hand does not sweep the
            # object sideways off the table on the way in.
            want[:, 2] += 0.06 * (1.0 - i / max(1, reach_end))
            grip = -1.0
        elif i < close_end:
            grip = -1.0 + 2.0 * (i - reach_end) / max(1, close_end - reach_end)
        else:
            lift = (i - close_end) / max(1, n - close_end)
            want[:, 2] += 0.18 * lift
            grip = 1.0

        a = torch.clamp((want - center) / half, -1.0, 1.0)
        action = torch.cat([a, torch.full_like(a[:, :1], grip)], dim=1)
        obs_dict, rew, term, trunc, _ = env.step(action)
        obs = obs_dict["policy"]

        root_z = robot.data.root_pos_w[:, 2]
        min_root_z = min(min_root_z, float(root_z.min()))
        if args_cli.video:
            fr = _rgb(inner.render())
            if fr is not None:
                frames.append(fr)
        if i % 10 == 0 or i == n - 1:
            err = torch.norm(obs[:, _DELTA], dim=1).mean().item()
            lifted = (obj.data.root_pos_w[:, 2] - rest_z).mean().item()
            log.append(
                (
                    i,
                    float(root_z.min()),
                    err,
                    lifted,
                    float(rew.mean()),
                    float((term | trunc).float().sum()),
                )
            )

    print(f"\n{'step':>5} {'min root_z':>11} {'hand-obj':>9} {'lift vs rest':>13} {'reward':>8} {'done':>5}")
    for row in log:
        print(f"{row[0]:5d} {row[1]:11.3f} {row[2]:9.3f} {row[3]:13.3f} {row[4]:8.3f} {row[5]:5.0f}")

    final_lift = (obj.data.root_pos_w[:, 2] - rest_z).detach().cpu().numpy()
    print(f"\n[smoke] min root_z over the run: {min_root_z:.3f}  (fell if < 0.5)")
    print(f"[smoke] final lift above rest, per env: {np.round(final_lift, 3).tolist()}")
    print(f"[smoke] picked up (>3 cm above rest): {int((final_lift > 0.03).sum())}/{len(final_lift)}")

    if args_cli.video and frames:
        out = args_cli.video.resolve()
        out.parent.mkdir(parents=True, exist_ok=True)
        import imageio.v2 as imageio

        imageio.mimsave(str(out), frames, fps=float(args_cli.fps), codec="libx264", pixelformat="yuv420p")
        print(f"[smoke] wrote {out} ({len(frames)} frames)")
    env.close()


if __name__ == "__main__":
    # os._exit in a bare finally would kill the process before the traceback
    # reaches the log, turning any failure into a silent early exit.
    code = 0
    try:
        main()
    except Exception:
        traceback.print_exc()
        sys.stdout.flush()
        sys.stderr.flush()
        code = 1
    finally:
        simulation_app.close()
    os._exit(code)

	#!/usr/bin/env python3
	"""Drive V2D-G1-SonicManip with a scripted pick, before any training.

	A learned policy failing tells you nothing about whether the environment is
	sound. This runs an open-loop reach / close / lift schedule instead, so the
	question is narrow: with a sensible action sequence, does the robot stay up,
	does SONIC put the hand where it was told, and does the object come off the
	table? Anything that fails here would have failed silently during training.

	cd simulation
	./run_isaaclab.sh --python scripts/smoke_sonic_manip.py --headless
	"""

	from __future__ import annotations

	import argparse
	import os
	import sys
	import traceback
	from pathlib import Path

	from isaaclab.app import AppLauncher

	if hasattr(sys.stdout, "reconfigure"):
	sys.stdout.reconfigure(line_buffering=True)

	parser = argparse.ArgumentParser(description="Scripted-pick smoke test for the SONIC manip env.")
	parser.add_argument("--task", type=str, default="V2D-G1-SonicManip-Play-v0")
	parser.add_argument("--num_envs", type=int, default=4)
	parser.add_argument("--steps", type=int, default=150)
	parser.add_argument("--video", type=Path, default=None, help="optional MP4 of env 0")
	parser.add_argument("--fps", type=float, default=25.0)
	AppLauncher.add_app_launcher_args(parser)
	args_cli = parser.parse_args()
	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 numpy as np # noqa: E402
	import torch # noqa: E402
	from isaaclab.utils.math import quat_apply # noqa: E402
	from isaaclab_tasks.utils import parse_env_cfg # noqa: E402

	import v2d_sim # noqa: E402, F401
	from v2d_sim.tasks.g1_sonic_manip.mdp.observations import t as _t # noqa: E402

	# Observation layout from PolicyObsCfg, in declaration order.
	_GRAV, _WRIST, _OBJ, _DELTA, _OBJV, _GRIP = (
	slice(0, 3),
	slice(3, 6),
	slice(6, 9),
	slice(9, 12),
	slice(12, 15),
	slice(15, 16),
	)


	#: Long enough for a marginally balanced object to visibly roll or topple.
	_SETTLE_STEPS = 20


	def _object_repose(obj) -> tuple[torch.Tensor, torch.Tensor]:
	"""Object position and how far its up axis leans off vertical, in degrees.

	The can's up axis is body +y: the reconstructed mesh is y-up, which is the
	whole reason its spawn rotation matters.
	"""
	pos = _t(obj.data.root_pos_w).clone()
	quat = _t(obj.data.root_quat_w)
	body_up = torch.tensor([0.0, 1.0, 0.0], device=quat.device).expand(quat.shape[0], 3)
	up_w = quat_apply(quat, body_up)
	return pos, torch.rad2deg(torch.arccos(up_w[:, 2].clamp(-1.0, 1.0)))


	def _rgb(img) -> np.ndarray \| None:
	if img is None:
	return None
	if hasattr(img, "detach"):
	img = img.detach().cpu().numpy()
	img = np.asarray(img)
	if img.ndim == 4:
	img = img[0]
	img = img[..., :3]
	if img.dtype != np.uint8:
	scale = 255.0 if float(np.nanmax(img)) <= 1.5 else 1.0
	img = np.clip(img * scale, 0, 255).astype(np.uint8)
	h, w = img.shape[:2]
	return img[: h - (h % 2), : w - (w % 2)]


	def main() -> None:
	env_cfg = parse_env_cfg(args_cli.task, device=args_cli.device, num_envs=args_cli.num_envs)
	render = "rgb_array" if args_cli.video else None
	env = gym.make(args_cli.task, cfg=env_cfg, render_mode=render)
	inner = env.unwrapped
	obs_dict, _ = env.reset()
	obs = obs_dict["policy"]

	acfg = env_cfg.actions.wrist
	lo = torch.tensor(
	[acfg.wrist_x[0], acfg.wrist_y[0], acfg.wrist_z[0]], device=inner.device
	)
	hi = torch.tensor(
	[acfg.wrist_x[1], acfg.wrist_y[1], acfg.wrist_z[1]], device=inner.device
	)
	center, half = (lo + hi) / 2.0, (hi - lo) / 2.0

	robot = inner.scene["robot"]
	obj = inner.scene["object"]
	table_top = env_cfg.scene.table.init_state.pos[2] + 0.5 * env_cfg.scene.table.spawn.size[2]
	# Height gained is measured from the resting pose, not the table top: the
	# object origin already sits half an object-height above the table.
	rest_z = env_cfg.scene.object.init_state.pos[2]

	print(f"[smoke] task={args_cli.task} envs={inner.num_envs} device={inner.device}")
	print(f"[smoke] obs dim={obs.shape[1]} action dim={inner.action_manager.total_action_dim}")
	print(f"[smoke] joints={len(robot.joint_names)} table_top_z={table_top:.3f}")
	print(f"[smoke] object rests at z={rest_z:.3f} ({rest_z - table_top:.3f} above the table)")
	print(f"[smoke] wrist box x{tuple(acfg.wrist_x)} y{tuple(acfg.wrist_y)} z{tuple(acfg.wrist_z)}")
	print(f"[smoke] object start (pelvis frame) = {obs[0, _OBJ].tolist()}")

	# Does the can sit still before anything touches it? A wrong spawn rotation
	# leaves it balanced on a curved side, which the run summary cannot
	# distinguish from a grasp that knocked it over.
	p0, tilt0 = _object_repose(obj)
	hold = torch.zeros(inner.num_envs, 4, device=inner.device)
	hold[:, 2] = 1.0 # wrist parked at the top of its box, clear of the object
	hold[:, 3] = -1.0
	for _ in range(_SETTLE_STEPS):
	env.step(hold)
	p1, tilt1 = _object_repose(obj)
	drift = torch.norm((p1 - p0)[:, :2], dim=1)
	print(
	f"[smoke] untouched for {_SETTLE_STEPS} steps: xy drift {drift.max():.4f} m, "
	f"tilt {tilt0.max():.1f} -> {tilt1.max():.1f} deg from vertical "
	"(firm if drift < ~5 mm and tilt stays put)"
	)
	obs_dict, _ = env.reset()
	obs = obs_dict["policy"]

	n = args_cli.steps
	reach_end, close_end = int(0.40 * n), int(0.55 * n)
	frames: list[np.ndarray] = []
	log: list[tuple[int, float, float, float, float, float]] = []
	min_root_z = float("inf")

	for i in range(n):
	if not simulation_app.is_running():
	break
	object_b = obs[:, _OBJ]
	want = object_b.clone()
	if i < reach_end:
	# Approach from slightly above, so the hand does not sweep the
	# object sideways off the table on the way in.
	want[:, 2] += 0.06 * (1.0 - i / max(1, reach_end))
	grip = -1.0
	elif i < close_end:
	grip = -1.0 + 2.0 * (i - reach_end) / max(1, close_end - reach_end)
	else:
	lift = (i - close_end) / max(1, n - close_end)
	want[:, 2] += 0.18 * lift
	grip = 1.0

	a = torch.clamp((want - center) / half, -1.0, 1.0)
	action = torch.cat([a, torch.full_like(a[:, :1], grip)], dim=1)
	obs_dict, rew, term, trunc, _ = env.step(action)
	obs = obs_dict["policy"]

	root_z = robot.data.root_pos_w[:, 2]
	min_root_z = min(min_root_z, float(root_z.min()))
	if args_cli.video:
	fr = _rgb(inner.render())
	if fr is not None:
	frames.append(fr)
	if i % 10 == 0 or i == n - 1:
	err = torch.norm(obs[:, _DELTA], dim=1).mean().item()
	lifted = (obj.data.root_pos_w[:, 2] - rest_z).mean().item()
	log.append(
	(
	i,
	float(root_z.min()),
	err,
	lifted,
	float(rew.mean()),
	float((term \| trunc).float().sum()),
	)
	)

	print(f"\n{'step':>5} {'min root_z':>11} {'hand-obj':>9} {'lift vs rest':>13} {'reward':>8} {'done':>5}")
	for row in log:
	print(f"{row[0]:5d} {row[1]:11.3f} {row[2]:9.3f} {row[3]:13.3f} {row[4]:8.3f} {row[5]:5.0f}")

	final_lift = (obj.data.root_pos_w[:, 2] - rest_z).detach().cpu().numpy()
	print(f"\n[smoke] min root_z over the run: {min_root_z:.3f} (fell if < 0.5)")
	print(f"[smoke] final lift above rest, per env: {np.round(final_lift, 3).tolist()}")
	print(f"[smoke] picked up (>3 cm above rest): {int((final_lift > 0.03).sum())}/{len(final_lift)}")

	if args_cli.video and frames:
	out = args_cli.video.resolve()
	out.parent.mkdir(parents=True, exist_ok=True)
	import imageio.v2 as imageio

	imageio.mimsave(str(out), frames, fps=float(args_cli.fps), codec="libx264", pixelformat="yuv420p")
	print(f"[smoke] wrote {out} ({len(frames)} frames)")
	env.close()


	if __name__ == "__main__":
	# os._exit in a bare finally would kill the process before the traceback
	# reaches the log, turning any failure into a silent early exit.
	code = 0
	try:
	main()
	except Exception:
	traceback.print_exc()
	sys.stdout.flush()
	sys.stderr.flush()
	code = 1
	finally:
	simulation_app.close()
	os._exit(code)