536 perturbed collection patch v4
Browse files
debug/536/perturb_v4/README.md
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Perturbed collection patch v4 (536 / nova, 2026-11-15)
|
| 2 |
+
Base: MagicSim mobile-ready 7fc2f64009 (any tree with src/magicsim/Env/Environment/SyncRobotEnv.py of that shape).
|
| 3 |
+
1. `python apply_patch.py <repo root>`: copies `perturb.py` to `src/magicsim/Collect/perturb.py` and hooks it into `SyncRobotEnv.step` (collect path only:
|
| 4 |
+
the recorded L3 `robot_action` stays the expert's; the robot executes it + a temporary offset) and `reset_idx`.
|
| 5 |
+
2. Enable per kit: `export MMB_COLLECT_PERTURB=1 MMB_PERTURB_SEED=<episode seed>` before the usual collect command (defaults in perturb.py docstring:
|
| 6 |
+
bursts p=0.01/tick, 25-75 ticks, warmup 100; base vx/vy 0.03 m/s, wz 0.06 rad/s only while the expert drives (|cmd| > 0.02); EEF xyz 0.008 m, bursts 15-40 ticks, no EEF push while the gripper is commanded closed; v4: EEF push scaled by the expert EEF speed (down to 15% in slow final alignment)).
|
| 7 |
+
The run log prints `[perturb] ON ...` and `[perturb] env 0: N bursts in T ticks`.
|
| 8 |
+
3. HF destination (owner): `<task>/<robot>/dr_on_perturb|dr_off_perturb/seed_<n>/` (dr_on/off = visual DR half, as the existing folders). 536 seeds 9001+, 850 seeds 9501+.
|
| 9 |
+
Smoke v2 on 536: N033 go2 4/4 kept (ALL PASS); N010 sawyer 0/3 (violation:furniture_overlap); N025 stretch3 1 fail of 1 so far.
|
debug/536/perturb_v4/apply_patch.py
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""apply_patch.py <repo root>: install magicsim/Collect/perturb.py and hook it into SyncRobotEnv.step / reset_idx (collect path only; the L3 replay / eval paths are untouched)."""
|
| 2 |
+
import shutil, sys
|
| 3 |
+
from pathlib import Path
|
| 4 |
+
root = Path(sys.argv[1]); here = Path(__file__).resolve().parent
|
| 5 |
+
shutil.copy(here / "perturb.py", root / "src/magicsim/Collect/perturb.py")
|
| 6 |
+
p = root / "src/magicsim/Env/Environment/SyncRobotEnv.py"; s = p.read_text()
|
| 7 |
+
a = " self.direct_robot_action = False\n"
|
| 8 |
+
a2 = a + " self._l3_perturb = None # Collect.perturb.L3Perturber (MMB_COLLECT_PERTURB=1), built on the first collect-path step\n self._l3_perturb_checked = False\n"
|
| 9 |
+
b = """ step_info = self.robot_manager.step(
|
| 10 |
+
action=action_info["robot_action"], env_ids=env_ids
|
| 11 |
+
)"""
|
| 12 |
+
b2 = """ exec_action = action_info["robot_action"]
|
| 13 |
+
if not self.direct_robot_action:
|
| 14 |
+
# perturbed collection (MMB_COLLECT_PERTURB=1): the recorded L3 stays the expert's; the robot executes it plus a temporary offset (Collect/perturb.py)
|
| 15 |
+
if not self._l3_perturb_checked:
|
| 16 |
+
self._l3_perturb_checked = True
|
| 17 |
+
from magicsim.Collect.perturb import L3Perturber
|
| 18 |
+
rm = self.robot_manager
|
| 19 |
+
hint = ",".join([str(k) for k in (getattr(rm, "robots", None) or {})] + [type(v).__name__ for v in (getattr(rm, "robots", None) or {}).values()])
|
| 20 |
+
self._l3_perturb = L3Perturber.from_env(self.num_envs, int(getattr(rm, "total_action_dim", exec_action.shape[-1])), hint)
|
| 21 |
+
if self._l3_perturb is not None:
|
| 22 |
+
exec_action = self._l3_perturb.apply(exec_action, env_ids)
|
| 23 |
+
step_info = self.robot_manager.step(
|
| 24 |
+
action=exec_action, env_ids=env_ids
|
| 25 |
+
)"""
|
| 26 |
+
c = " super().reset_idx(env_ids=env_id_list, seed=seed, options=options)\n"
|
| 27 |
+
c2 = c + " if getattr(self, \"_l3_perturb\", None) is not None:\n self._l3_perturb.reset(env_id_list)\n"
|
| 28 |
+
for x in (a, b, c):
|
| 29 |
+
assert s.count(x) == 1, ("pattern count", s.count(x), x[:60])
|
| 30 |
+
s = s.replace(a, a2).replace(b, b2).replace(c, c2); p.write_text(s); print("patched", p)
|
debug/536/perturb_v4/perturb.py
ADDED
|
@@ -0,0 +1,119 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Perturbed collection (DART-style), owner 2026-11-15 22:30 CST: "把加扰动采数据这一点带上".
|
| 2 |
+
|
| 3 |
+
The expert's L3 ``robot_action`` (PlannerManager output) is RECORDED unchanged; the robot EXECUTES it plus a temporary offset. The Collect layer is closed loop (it
|
| 4 |
+
re-reads the robot / object state every tick), so after (and during) an offset it drives the robot back: the dataset then holds "off the nominal path -> the expert's
|
| 5 |
+
correction" states, which plain expert data never visits and which a policy needs to recover from its own drift.
|
| 6 |
+
|
| 7 |
+
Off unless ``MMB_COLLECT_PERTURB=1``. Per env, a burst starts with probability ``P`` per control tick (after ``WARMUP`` ticks from the reset, none while one is active);
|
| 8 |
+
a burst lasts ``LEN_MIN..LEN_MAX`` ticks with a constant offset (linear ramp in / out over ``RAMP`` ticks) on
|
| 9 |
+
* the base velocity command (vx, vy, wz): N(0, BASE_V) m/s, N(0, BASE_W) rad/s (stretch3 has no vy: 0), applied only on ticks where the expert itself drives
|
| 10 |
+
the base (max |vx, vy, wz| of its command > BASE_GATE): a parked base (manipulating at a table / rack) is never pushed (v1 pushed a parked Sawyer into the tool rack:
|
| 11 |
+
2 of 2 N010 runs ended in violation:furniture_overlap),
|
| 12 |
+
* the EEF target position (x, y, z) : N(0, EEF_M) m per axis, only while the expert commands the gripper OPEN (v3, 2026-11-16: an offset while the
|
| 13 |
+
gripper is closed = carrying / inserting jammed the N025 insertions and broke grasp servos: N025 sawyer 0/3 kept vs 97 % unperturbed).
|
| 14 |
+
The gripper, the base height and the EEF rotation are never touched; NaN entries (= "not commanded this tick") stay NaN.
|
| 15 |
+
v4 (owner 2026-11-16 00:56 CST "越到接近插桩的时候,扰动就越小"): the EEF offset is scaled by the expert's own EEF target speed, clip(|dp| / SPEED_REF, SCALE_MIN, 1)
|
| 16 |
+
per tick (|dp| = change of the expert EEF target position since the last tick, smoothed over ~10 ticks): fast reaching moves get the full offset, the slow final
|
| 17 |
+
alignment / insertion approach gets down to SCALE_MIN of it. MMB_PERTURB_SPEED_REF [0.004 m/tick = 0.2 m/s], MMB_PERTURB_SCALE_MIN [0.15].
|
| 18 |
+
Environment (defaults in brackets): MMB_PERTURB_P [0.01], MMB_PERTURB_LEN_MIN [15], MMB_PERTURB_LEN_MAX [40], MMB_PERTURB_RAMP [5], MMB_PERTURB_WARMUP [100],
|
| 19 |
+
MMB_PERTURB_BASE_V [0.03], MMB_PERTURB_BASE_W [0.06], MMB_PERTURB_BASE_GATE [0.02], MMB_PERTURB_EEF_M [0.008], MMB_PERTURB_SEED [none].
|
| 20 |
+
L3 layout (canonical_action.join_raw): Go2 [vx vy wz h | x y z | qw qx qy qz | grip] (12), Sawyer / Stretch [vx vy wz | x y z | qw qx qy qz | grip] (11)."""
|
| 21 |
+
|
| 22 |
+
from __future__ import annotations
|
| 23 |
+
|
| 24 |
+
import os
|
| 25 |
+
from typing import Optional, Sequence
|
| 26 |
+
|
| 27 |
+
import numpy as np
|
| 28 |
+
import torch
|
| 29 |
+
|
| 30 |
+
ENV = "MMB_COLLECT_PERTURB"
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
def _f(name: str, default: float) -> float:
|
| 34 |
+
v = os.environ.get(f"MMB_PERTURB_{name}", "").strip()
|
| 35 |
+
return float(v) if v else default
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
class L3Perturber:
|
| 39 |
+
def __init__(self, num_envs: int, action_dim: int, robot_hint: str = ""):
|
| 40 |
+
self.n, self.D = int(num_envs), int(action_dim)
|
| 41 |
+
hint = robot_hint.lower()
|
| 42 |
+
self.go2 = self.D == 12 or "go2" in hint
|
| 43 |
+
self.no_vy = ("hello" in hint or "stretch" in hint) and not self.go2
|
| 44 |
+
self.pos = slice(4, 7) if self.go2 else slice(3, 6)
|
| 45 |
+
self.p, self.lmin, self.lmax = _f("P", 0.01), int(_f("LEN_MIN", 15)), int(_f("LEN_MAX", 40))
|
| 46 |
+
self.ramp, self.warmup = max(1, int(_f("RAMP", 5))), int(_f("WARMUP", 100))
|
| 47 |
+
self.sv, self.sw, self.se = _f("BASE_V", 0.03), _f("BASE_W", 0.06), _f("EEF_M", 0.008)
|
| 48 |
+
self.gate = _f("BASE_GATE", 0.02)
|
| 49 |
+
self.vref, self.smin = _f("SPEED_REF", 0.004), _f("SCALE_MIN", 0.15)
|
| 50 |
+
self.prev_p = np.full((self.n, 3), np.nan, np.float32); self.vel = np.zeros(self.n, np.float32)
|
| 51 |
+
seed = os.environ.get("MMB_PERTURB_SEED", "").strip()
|
| 52 |
+
self.rng = np.random.default_rng(int(seed) if seed else None)
|
| 53 |
+
self.tick = np.zeros(self.n, np.int64) # ticks since the env's last reset
|
| 54 |
+
self.left = np.zeros(self.n, np.int64) # ticks left in the current burst (0 = none)
|
| 55 |
+
self.length = np.ones(self.n, np.int64)
|
| 56 |
+
self.off = np.zeros((self.n, self.D), np.float32)
|
| 57 |
+
self.bursts = np.zeros(self.n, np.int64) # bursts in the current episode (for the log)
|
| 58 |
+
print(f"[perturb] ON: robot={'go2' if self.go2 else ('stretch3' if self.no_vy else 'sawyer/other')} D={self.D} p={self.p} len={self.lmin}-{self.lmax} "
|
| 59 |
+
f"warmup={self.warmup} base_v={self.sv} base_w={self.sw} gate={self.gate} eef_m={self.se} speed_ref={self.vref} scale_min={self.smin}", flush=True)
|
| 60 |
+
|
| 61 |
+
@classmethod
|
| 62 |
+
def from_env(cls, num_envs: int, action_dim: int, robot_hint: str = "") -> Optional["L3Perturber"]:
|
| 63 |
+
if os.environ.get(ENV, "").strip() != "1":
|
| 64 |
+
return None
|
| 65 |
+
return cls(num_envs, action_dim, robot_hint)
|
| 66 |
+
|
| 67 |
+
def reset(self, env_ids: Sequence[int]) -> None:
|
| 68 |
+
for e in env_ids:
|
| 69 |
+
e = int(e)
|
| 70 |
+
if 0 <= e < self.n:
|
| 71 |
+
if self.bursts[e]:
|
| 72 |
+
print(f"[perturb] env {e}: {int(self.bursts[e])} bursts in {int(self.tick[e])} ticks", flush=True)
|
| 73 |
+
self.tick[e] = 0; self.left[e] = 0; self.bursts[e] = 0; self.off[e] = 0; self.prev_p[e] = np.nan; self.vel[e] = 0
|
| 74 |
+
|
| 75 |
+
def _new_offset(self) -> np.ndarray:
|
| 76 |
+
o = np.zeros(self.D, np.float32)
|
| 77 |
+
o[0], o[1], o[2] = self.rng.normal(0, self.sv), self.rng.normal(0, self.sv), self.rng.normal(0, self.sw)
|
| 78 |
+
if self.no_vy:
|
| 79 |
+
o[1] = 0.0
|
| 80 |
+
o[self.pos] = self.rng.normal(0, self.se, 3)
|
| 81 |
+
return o
|
| 82 |
+
|
| 83 |
+
def apply(self, action: torch.Tensor, env_ids: Sequence[int]) -> torch.Tensor:
|
| 84 |
+
"""The tensor to EXECUTE: a copy of ``action`` with the active offsets added (``action`` itself is not modified: it is what gets recorded)."""
|
| 85 |
+
if not isinstance(action, torch.Tensor) or action.dim() != 2 or action.shape[-1] != self.D:
|
| 86 |
+
return action
|
| 87 |
+
add = np.zeros((action.shape[0], self.D), np.float32)
|
| 88 |
+
for e in env_ids:
|
| 89 |
+
e = int(e)
|
| 90 |
+
if not (0 <= e < self.n) or e >= action.shape[0]:
|
| 91 |
+
continue
|
| 92 |
+
self.tick[e] += 1
|
| 93 |
+
p_now = action[e, self.pos].detach().float().cpu().numpy() # expert EEF target position (recorded L3)
|
| 94 |
+
if np.isfinite(p_now).all():
|
| 95 |
+
if np.isfinite(self.prev_p[e]).all():
|
| 96 |
+
self.vel[e] = 0.9 * self.vel[e] + 0.1 * float(np.linalg.norm(p_now - self.prev_p[e]))
|
| 97 |
+
self.prev_p[e] = p_now
|
| 98 |
+
if self.left[e] == 0 and self.tick[e] > self.warmup and self.rng.random() < self.p:
|
| 99 |
+
self.length[e] = self.left[e] = int(self.rng.integers(self.lmin, self.lmax + 1))
|
| 100 |
+
self.off[e] = self._new_offset(); self.bursts[e] += 1
|
| 101 |
+
if self.left[e] > 0:
|
| 102 |
+
k = self.length[e] - self.left[e] # ticks into the burst
|
| 103 |
+
g = min(1.0, (k + 1) / self.ramp, self.left[e] / self.ramp)
|
| 104 |
+
add[e] = self.off[e] * g
|
| 105 |
+
cmd = action[e, 0:3].detach().float().cpu().numpy()
|
| 106 |
+
if not np.isfinite(cmd).all() or np.abs(cmd).max() <= self.gate: # parked base: no base push
|
| 107 |
+
add[e, 0:3] = 0.0
|
| 108 |
+
add[e, self.pos] *= min(1.0, max(self.smin, self.vel[e] / self.vref)) # v4: slow expert = fine alignment / insertion -> small push
|
| 109 |
+
grip = float(action[e, -1]) # the gripper command is the last L3 column (closure, > 0.5 = close)
|
| 110 |
+
if np.isfinite(grip) and grip > 0.5: # gripper closed (holding / inserting): no EEF push
|
| 111 |
+
add[e, self.pos] = 0.0
|
| 112 |
+
self.left[e] -= 1
|
| 113 |
+
if not add.any():
|
| 114 |
+
return action
|
| 115 |
+
t = torch.as_tensor(add, dtype=action.dtype, device=action.device)
|
| 116 |
+
out = action.clone()
|
| 117 |
+
fin = torch.isfinite(out)
|
| 118 |
+
out[fin] = out[fin] + t[fin] # NaN = not commanded: stays NaN
|
| 119 |
+
return out
|