simulation/scripts/ablate_actions.py

#!/usr/bin/env python3
"""Open-loop action ablation on V2D-G1-SonicManip, before any PPO.

Inject a named action stream (zeros, scripted pick, right-to-left hover sweep,
relocated clip). Rank streams on physics: did the robot stay up, did SONIC track
the wrist command, did the object move, how jerky was the command. Do not rank
on resemblance to HaWoR.

  cd simulation
  ./run_isaaclab.sh --python scripts/ablate_actions.py --headless \\
      --policies zero,scripted_4d,rel_hover,raw_clip --video-dir runs/action_ablate

``vae0`` / ``vae_grasp`` are reserved; the env is still 4-D and those names
error out until the hand prior is wired.
"""

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="Open-loop action ablation for SonicManip.")
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(
    "--policies",
    type=str,
    default="zero,scripted_4d,rel_hover,raw_clip",
    help="comma-separated names, or 'all' (excludes VAE stubs)",
)
parser.add_argument(
    "--ema",
    type=float,
    default=1.0,
    help="causal EMA on the 4-D command; 1 = off. *_smooth policies default to 0.3",
)
parser.add_argument(
    "--hover",
    type=float,
    nargs=3,
    default=(0.0, 0.0, 0.02),
    metavar=("X", "Y", "Z"),
    help="pelvis-frame offset on the frozen object pose (rel_hover sweep endpoint)",
)
parser.add_argument(
    "--contact-dist",
    type=float,
    default=0.05,
    help="rel_hover stops the y-sweep when wrist-object is closer than this (m)",
)
parser.add_argument(
    "--sweep-frac",
    type=float,
    default=0.75,
    help="unused for rel_hover (kept for CLI compat); speed is --sweep-speed",
)
parser.add_argument(
    "--sweep-speed",
    type=float,
    default=0.08,
    help="rel_hover max wrist-target speed in m/s (pelvis). 0.08 ≈ 8 cm/s",
)
parser.add_argument(
    "--sweep-margin",
    type=float,
    default=0.06,
    help="start the lateral sweep this many metres to the robot's right of the object",
)
parser.add_argument("--clip", type=Path, default=None, help="isaaclab_replay.npz")
parser.add_argument("--wuji", type=Path, default=None, help="g1_wuji_retarget.npz for clip grip")
parser.add_argument("--video-dir", type=Path, default=None)
parser.add_argument("--log-dir", type=Path, default=None)
parser.add_argument("--fps", type=float, default=25.0)
parser.add_argument("--jitter", action="store_true", help="keep object xy/yaw reset jitter")
parser.add_argument("--keep-term", action="store_true", help="keep fall / object-off-table resets")
parser.add_argument("--list-policies", action="store_true")
AppLauncher.add_app_launcher_args(parser)
args_cli = parser.parse_args()

# Names mirrored from openloop.POLICY_NAMES so --list works before Isaac boots.
_POLICY_NAMES = (
    "zero",
    "park",
    "scripted_4d",
    "raw_clip",
    "smooth_clip",
    "rel_hover",
    "rel_hover_smooth",
    "vae0",
    "vae_grasp",
)

if args_cli.list_policies:
    print("\n".join(_POLICY_NAMES))
    sys.exit(0)

if args_cli.video_dir:
    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_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
from v2d_sim.tasks.g1_sonic_manip.openloop import (  # noqa: E402
    DEFAULT_REPLAY_NPZ,
    DEFAULT_WUJI_NPZ,
    default_ema,
    expand_policies,
    is_vae_policy,
    load_clip_wrist_stream,
    scripted_grip,
    scripted_lift_z,
)

_GRAV, _WRIST, _OBJ, _DELTA, _OBJV, _GRIP = (
    slice(0, 3),
    slice(3, 6),
    slice(6, 9),
    slice(9, 12),
    slice(12, 15),
    slice(15, 16),
)
_SETTLE_STEPS = 20


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)]


class Ema4:
    def __init__(self, alpha: float) -> None:
        self.alpha = float(alpha)
        self.state: torch.Tensor | None = None

    def reset(self) -> None:
        self.state = None

    def __call__(self, a: torch.Tensor) -> torch.Tensor:
        if self.alpha >= 1.0 - 1e-6:
            return a
        if self.state is None:
            self.state = a.clone()
        else:
            self.state = self.alpha * a + (1.0 - self.alpha) * self.state
        return self.state


def _box(env_cfg, device) -> tuple[torch.Tensor, torch.Tensor, np.ndarray, np.ndarray]:
    acfg = env_cfg.actions.wrist
    lo = torch.tensor([acfg.wrist_x[0], acfg.wrist_y[0], acfg.wrist_z[0]], device=device)
    hi = torch.tensor([acfg.wrist_x[1], acfg.wrist_y[1], acfg.wrist_z[1]], device=device)
    center, half = (lo + hi) / 2.0, (hi - lo) / 2.0
    return center, half, center.detach().cpu().numpy(), half.detach().cpu().numpy()


def _pack_action(xyz_b: torch.Tensor, grip: torch.Tensor | float, center, half) -> torch.Tensor:
    a = torch.clamp((xyz_b - center) / half, -1.0, 1.0)
    if not torch.is_tensor(grip):
        g = torch.full((a.shape[0], 1), float(grip), device=a.device, dtype=a.dtype)
    else:
        g = grip.reshape(a.shape[0], 1).to(device=a.device, dtype=a.dtype)
    return torch.cat([a, g], dim=1)


class ActionStream:
    def __init__(
        self,
        name: str,
        *,
        center,
        half,
        center_np,
        half_np,
        hover_obj: torch.Tensor,
        clip,
        n_steps: int,
        dt: float = 0.04,
    ) -> None:
        self.name = name
        self.center, self.half = center, half
        self.center_np, self.half_np = center_np, half_np
        self.hover_obj = hover_obj
        self.clip = clip
        self.n_steps = n_steps
        self.dt = float(dt)
        self._goal: torch.Tensor | None = None
        self._start: torch.Tensor | None = None
        self._cmd: torch.Tensor | None = None
        self._obj_xy0: torch.Tensor | None = None
        self._hit: torch.Tensor | None = None
        self._at_start: torch.Tensor | None = None

    def _rel_hover_sweep(self, i: int, obs: torch.Tensor) -> torch.Tensor:
        """Rate-limited approach from slightly right of the can; no lift.

        Parking at the wrist-box y-min first is ~12 cm off the object; at a
        few cm/s that leg eats the episode and the hand never arrives. Crawl
        from the achieved wrist to a point ``--sweep-margin`` to the right of
        the frozen object (clamped to the SONIC box), then slide +y onto the
        can. Contact is ignored until that approach point.
        """
        n_env, device = obs.shape[0], obs.device
        obj = obs[:, _OBJ]
        vmax = float(args_cli.sweep_speed)
        max_step = vmax * self.dt
        lo = self.center - self.half
        hi = self.center + self.half
        if self._goal is None:
            offset = self.hover_obj.to(device=device, dtype=obj.dtype).view(1, 3)
            self._goal = torch.minimum(hi, torch.maximum(lo, obj + offset))
            y_right = torch.clamp(
                self._goal[:, 1] - float(args_cli.sweep_margin), min=float(lo[1])
            )
            self._start = self._goal.clone()
            self._start[:, 1] = y_right
            self._cmd = torch.minimum(hi, torch.maximum(lo, obs[:, _WRIST].clone()))
            self._obj_xy0 = obj[:, :2].clone()
            self._hit = torch.zeros(n_env, dtype=torch.bool, device=device)
            already = obs[:, _WRIST][:, 1] <= self._start[:, 1] + 0.01
            self._at_start = already
            path = torch.norm(self._start - self._cmd, dim=1) + torch.norm(
                self._goal - self._start, dim=1
            )
            eta = float(path.mean() / max(vmax, 1e-6))
            print(
                f"  sweep {vmax:.3f} m/s  margin={args_cli.sweep_margin:.3f} m  "
                f"path≈{float(path.mean()):.3f} m  eta≈{eta:.1f}s  "
                f"goal_b={self._goal[0].detach().cpu().tolist()}  "
                f"(need ~{int(eta / self.dt) + 1} steps; this run has {self.n_steps})"
            )

        target = torch.where(self._at_start.unsqueeze(-1), self._goal, self._start)
        delta = target - self._cmd
        dist = torch.linalg.norm(delta, dim=1, keepdim=True).clamp_min(1e-8)
        step = torch.clamp(dist, max=max_step)
        moved = self._cmd + step * delta / dist
        live = (~self._hit).unsqueeze(-1)
        self._cmd = torch.where(live, moved, self._cmd)
        self._at_start |= (~self._hit) & (dist.squeeze(-1) <= max_step + 1e-4)

        dist_obj = torch.norm(obs[:, _DELTA], dim=1)
        shoved = torch.norm(obj[:, :2] - self._obj_xy0, dim=1) > 0.02
        self._hit |= self._at_start & ((dist_obj < args_cli.contact_dist) | shoved)
        grip = torch.where(
            self._hit,
            torch.ones(n_env, device=device, dtype=obj.dtype),
            -torch.ones(n_env, device=device, dtype=obj.dtype),
        )
        return _pack_action(self._cmd, grip, self.center, self.half)

    def raw(self, i: int, obs: torch.Tensor, robot, obj) -> torch.Tensor:
        n_env = obs.shape[0]
        device = obs.device
        name = self.name
        if name == "zero":
            return torch.zeros(n_env, 4, device=device)
        if name == "park":
            a = torch.zeros(n_env, 4, device=device)
            a[:, 2] = 1.0
            a[:, 3] = -1.0
            return a
        if name == "scripted_4d":
            want = obs[:, _OBJ].clone()
            reach_end = int(0.40 * self.n_steps)
            if i < reach_end:
                want[:, 2] += 0.06 * (1.0 - i / max(1, reach_end))
            else:
                want[:, 2] += scripted_lift_z(i, self.n_steps)
            return _pack_action(want, scripted_grip(i, self.n_steps), self.center, self.half)
        if name in ("rel_hover", "rel_hover_smooth"):
            return self._rel_hover_sweep(i, obs)
        if name in ("raw_clip", "smooth_clip"):
            if self.clip is None:
                raise RuntimeError("clip stream requested but isaaclab_replay.npz was not loaded")
            k = min(i, self.clip.wrist_b.shape[0] - 1)
            xyz = torch.tensor(self.clip.wrist_b[k], device=device, dtype=obs.dtype).expand(n_env, 3)
            g = float(self.clip.grip[k])
            return _pack_action(xyz, g, self.center, self.half)
        raise RuntimeError(f"unhandled policy {name}")


def _prepare_cfg(env_cfg):
    if not args_cli.jitter:
        env_cfg.events.reset_object.params["x_range"] = (0.0, 0.0)
        env_cfg.events.reset_object.params["y_range"] = (0.0, 0.0)
        env_cfg.events.reset_object.params["yaw_range"] = (0.0, 0.0)
    if not args_cli.keep_term:
        env_cfg.terminations.fallen = None
        env_cfg.terminations.object_fell = None
        env_cfg.episode_length_s = max(env_cfg.episode_length_s, args_cli.steps * 0.04 + 2.0)
    return env_cfg


def _rollout(env, inner, env_cfg, name: str, clip, log_dir: Path | None, video_dir: Path | None):
    if is_vae_policy(name):
        raise RuntimeError(
            f"{name}: SonicManip is still 4-D (wrist xyz + 1-D grip). "
            "Wire CoordEx kinematic_wrist_16k.pt before rolling out Δz=0 / grasp latent."
        )
    robot = inner.scene["robot"]
    obj = inner.scene["object"]
    rest_z = env_cfg.scene.object.init_state.pos[2]
    center, half, center_np, half_np = _box(env_cfg, inner.device)
    hover = torch.tensor(list(args_cli.hover), device=inner.device, dtype=torch.float32)
    alpha = default_ema(name, args_cli.ema)
    stream = ActionStream(
        name,
        center=center,
        half=half,
        center_np=center_np,
        half_np=half_np,
        hover_obj=hover,
        clip=clip,
        n_steps=args_cli.steps,
        dt=float(env_cfg.sim.dt * env_cfg.decimation),
    )
    ema = Ema4(alpha)

    obs_dict, _ = env.reset()
    hold = torch.zeros(inner.num_envs, 4, device=inner.device)
    hold[:, 2] = 1.0
    hold[:, 3] = -1.0
    for _ in range(_SETTLE_STEPS):
        env.step(hold)
    obs_dict, _ = env.reset()
    obs = obs_dict["policy"]
    ema.reset()

    n = args_cli.steps
    rec = {
        "action_raw": [],
        "action_ema": [],
        "wrist_cmd": [],
        "wrist_ach": [],
        "object_b": [],
        "delta": [],
        "root_z": [],
        "object_z": [],
        "reward": [],
        "done": [],
    }
    frames: list[np.ndarray] = []
    term = inner.action_manager.get_term("wrist")

    print(f"\n=== {name}  ema={alpha:.2f}  steps={n} ===")
    for i in range(n):
        if not simulation_app.is_running():
            break
        raw = stream.raw(i, obs, robot, obj)
        applied = ema(raw)
        obs_dict, rew, terminated, truncated, _ = env.step(applied)
        obs = obs_dict["policy"]
        rec["action_raw"].append(raw.detach().cpu().numpy())
        rec["action_ema"].append(applied.detach().cpu().numpy())
        rec["wrist_cmd"].append(term.wrist_target.detach().cpu().numpy())
        rec["wrist_ach"].append(obs[:, _WRIST].detach().cpu().numpy())
        rec["object_b"].append(obs[:, _OBJ].detach().cpu().numpy())
        rec["delta"].append(obs[:, _DELTA].detach().cpu().numpy())
        rec["root_z"].append(_t(robot.data.root_pos_w)[:, 2].detach().cpu().numpy())
        rec["object_z"].append(_t(obj.data.root_pos_w)[:, 2].detach().cpu().numpy())
        rec["reward"].append(rew.detach().cpu().numpy())
        rec["done"].append((terminated | truncated).float().detach().cpu().numpy())
        if video_dir is not None:
            fr = _rgb(inner.render())
            if fr is not None:
                frames.append(fr)

    stacked = {k: np.stack(v, axis=0) for k, v in rec.items()}
    d_act = np.linalg.norm(np.diff(stacked["action_ema"], axis=0), axis=-1)
    track = np.linalg.norm(stacked["wrist_ach"] - stacked["wrist_cmd"], axis=-1)
    hand_obj = np.linalg.norm(stacked["delta"], axis=-1)
    lift = stacked["object_z"] - rest_z
    summary = {
        "policy": name,
        "ema": alpha,
        "min_root_z": float(stacked["root_z"].min()),
        "mean_hand_obj": float(hand_obj.mean()),
        "final_hand_obj": float(hand_obj[-1].mean()),
        "max_lift": float(lift.max()),
        "final_lift": float(lift[-1].mean()),
        "picked": int((lift[-1] > 0.03).sum()),
        "n_env": int(lift.shape[1]),
        "mean_action_rate": float(d_act.mean()) if d_act.size else 0.0,
        "mean_track_err": float(track.mean()),
        "return": float(stacked["reward"].sum(axis=0).mean()),
        "n_done": float(stacked["done"].sum()),
    }
    print(
        f"  min_root_z={summary['min_root_z']:.3f}  "
        f"hand-obj mean/final={summary['mean_hand_obj']:.3f}/{summary['final_hand_obj']:.3f}  "
        f"lift max/final={summary['max_lift']:.3f}/{summary['final_lift']:.3f}  "
        f"picked={summary['picked']}/{summary['n_env']}  "
        f"|Δa|={summary['mean_action_rate']:.3f}  "
        f"track={summary['mean_track_err']:.3f}  "
        f"R={summary['return']:.2f}"
    )
    if stream._hit is not None:
        print(f"  contact {int(stream._hit.sum())}/{int(stream._hit.numel())} envs (sweep freeze)")

    tag = f"{name}_ema{alpha:.2f}".replace(".", "p")
    if log_dir is not None:
        log_dir.mkdir(parents=True, exist_ok=True)
        out = log_dir / f"{tag}.npz"
        scalars = {
            k: np.asarray(v) for k, v in summary.items() if k != "policy"
        }
        np.savez_compressed(out, **stacked, **{f"s_{k}": v for k, v in scalars.items()}, policy=np.array(name))
        print(f"  wrote {out}")
    if video_dir is not None and frames:
        video_dir.mkdir(parents=True, exist_ok=True)
        mp4 = video_dir / f"{tag}.mp4"
        import imageio.v2 as imageio

        imageio.mimsave(str(mp4), frames, fps=float(args_cli.fps), codec="libx264", pixelformat="yuv420p")
        print(f"  wrote {mp4} ({len(frames)} frames)")
    return summary


def main() -> None:
    names = expand_policies(args_cli.policies)
    env_cfg = parse_env_cfg(args_cli.task, device=args_cli.device, num_envs=args_cli.num_envs)
    env_cfg = _prepare_cfg(env_cfg)
    render = "rgb_array" if args_cli.video_dir else None
    env = gym.make(args_cli.task, cfg=env_cfg, render_mode=render)
    inner = env.unwrapped

    clip_path = args_cli.clip or DEFAULT_REPLAY_NPZ
    wuji_path = args_cli.wuji or DEFAULT_WUJI_NPZ
    need_clip = any(n in ("raw_clip", "smooth_clip") for n in names)
    clip_raw = clip_smooth = None
    if need_clip:
        if not Path(clip_path).is_file():
            raise FileNotFoundError(f"replay clip missing: {clip_path}")
        kwargs = dict(
            replay_npz=clip_path,
            wuji_npz=wuji_path if Path(wuji_path).is_file() else None,
            policy_fps=1.0 / (env_cfg.sim.dt * env_cfg.decimation),
            object_xy=(env_cfg.scene.object.init_state.pos[0], env_cfg.scene.object.init_state.pos[1]),
            table_z=env_cfg.scene.table.init_state.pos[2] + 0.5 * env_cfg.scene.table.spawn.size[2],
            pelvis_z=env_cfg.scene.robot.init_state.pos[2],
        )
        if any(n == "raw_clip" for n in names):
            clip_raw = load_clip_wrist_stream(**kwargs, smooth=False)
        if any(n == "smooth_clip" for n in names):
            clip_smooth = load_clip_wrist_stream(**kwargs, smooth=True)
        shown = clip_raw or clip_smooth
        acfg = env_cfg.actions.wrist
        lo = np.array([acfg.wrist_x[0], acfg.wrist_y[0], acfg.wrist_z[0]])
        hi = np.array([acfg.wrist_x[1], acfg.wrist_y[1], acfg.wrist_z[1]])
        inside = ((shown.wrist_b >= lo) & (shown.wrist_b <= hi)).all(axis=1).mean()
        print(
            f"[ablate] clip {clip_path}  T={shown.wrist_b.shape[0]}  "
            f"inside wrist box {100.0 * inside:.0f}%  "
            f"(frames outside are clamped — that is the result, not a loader bug)"
        )

    log_dir = args_cli.log_dir
    if log_dir is None and args_cli.video_dir is not None:
        log_dir = args_cli.video_dir
    if log_dir is None:
        log_dir = Path("runs/action_ablate")

    print(f"[ablate] task={args_cli.task}  envs={inner.num_envs}  device={inner.device}")
    print(f"[ablate] policies={names}  jitter={args_cli.jitter}  keep_term={args_cli.keep_term}")
    print(
        f"[ablate] hover_b={tuple(args_cli.hover)}  sweep_frac={args_cli.sweep_frac}  "
        f"contact={args_cli.contact_dist}  sweep_speed={args_cli.sweep_speed} m/s  log={log_dir}"
    )

    summaries = []
    for name in names:
        clip = clip_smooth if name == "smooth_clip" else clip_raw
        summaries.append(
            _rollout(env, inner, env_cfg, name, clip, log_dir, args_cli.video_dir)
        )

    print(f"\n{'policy':<20} {'ema':>5} {'root_z':>7} {'hand-obj':>8} {'lift':>7} {'|Δa|':>6} {'track':>6} {'R':>8}")
    for s in summaries:
        print(
            f"{s['policy']:<20} {s['ema']:5.2f} {s['min_root_z']:7.3f} "
            f"{s['final_hand_obj']:8.3f} {s['final_lift']:7.3f} "
            f"{s['mean_action_rate']:6.3f} {s['mean_track_err']:6.3f} {s['return']:8.2f}"
        )
    ranked = sorted(summaries, key=lambda s: (s["final_lift"], -s["final_hand_obj"]), reverse=True)
    print(f"\n[ablate] rank by final lift, then closer hand: {[s['policy'] for s in ranked]}")
    env.close()


if __name__ == "__main__":
    code = 0
    try:
        main()
    except Exception:
        traceback.print_exc()
        sys.stdout.flush()
        sys.stderr.flush()
        code = 1
    finally:
        simulation_app.close()
    os._exit(code)