Spaces:
Sleeping
Sleeping
File size: 7,624 Bytes
8c8ea52 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 | """
Domain randomization for the tabletop planning environment.
Randomizes everything that can vary in a real tabletop scene:
- number of objects
- which object is the target
- which bin is the target
- how many blockers, and what they block
- object positions (within reachable workspace)
- task instruction (generated from the sampled scene)
- distractor objects (present but irrelevant to the task)
- constraint type (fragile first, heavy last, etc.)
The model must generalize across all of this — not memorize one layout.
"""
import random
import string
from dataclasses import dataclass, field
from typing import Optional
OBJECT_NAMES = ["red_block", "blue_block", "green_block", "yellow_block", "purple_block"]
OBJECT_COLORS = {"red_block": "red", "blue_block": "blue", "green_block": "green",
"yellow_block": "yellow", "purple_block": "purple"}
BINS = ["A", "B"]
CONSTRAINTS = ["fragile_first", "heavy_last", "urgent_first", None, None, None] # None = no constraint
@dataclass
class ScenarioConfig:
# Objects actually present in the scene
objects: list[str] = field(default_factory=list)
# Which objects are targets (must be placed in a bin)
targets: dict[str, str] = field(default_factory=dict) # obj_name -> bin
# Blocking relationships: blocker -> blocked
blockers: dict[str, str] = field(default_factory=dict)
# Distractors: present but not part of the task
distractors: list[str] = field(default_factory=list)
# Active constraint
constraint: Optional[str] = None
# Generated instruction string
instruction: str = ""
# Object positions on the table (x, y) — workspace is roughly ±0.25
positions: dict[str, tuple] = field(default_factory=dict)
# Hidden traits, revealed via scan or proximity.
hidden_traits: dict[str, str] = field(default_factory=dict)
# Optional deadlines in steps for selected target objects.
deadlines: dict[str, int] = field(default_factory=dict)
def randomize_scenario(
n_objects: Optional[int] = None,
n_targets: Optional[int] = None,
n_blockers: Optional[int] = None,
force_blocked: bool = False,
scenario_pack: str = "default",
) -> ScenarioConfig:
"""
Generate a fully randomized scenario.
n_objects: total objects on table (default: random 2-5)
n_targets: how many must be placed in bins (default: random 1-2)
n_blockers: how many blocking relationships (default: random 0-2)
force_blocked: always have at least one blocker (good for training recovery)
"""
# Sample object count
pack = SCENARIO_PACKS.get(scenario_pack, OBJECT_NAMES)
total = n_objects or random.randint(2, 5)
total = min(total, len(pack))
# Pick which objects appear
present = random.sample(pack, total)
# Pick targets (subset of present objects)
max_targets = min(n_targets or random.randint(1, 2), len(present))
targets_list = random.sample(present, max_targets)
target_bins = {obj: random.choice(BINS) for obj in targets_list}
# Distractors = present but not targets
distractors = [o for o in present if o not in target_bins]
# Build blocking relationships
n_block = n_blockers if n_blockers is not None else random.randint(0, min(2, len(distractors)))
if force_blocked:
n_block = max(1, n_block)
blockers = {}
# A blocker must be a non-target (distractor) blocking a target
available_blockers = list(distractors)
available_targets = list(targets_list)
random.shuffle(available_blockers)
random.shuffle(available_targets)
for i in range(min(n_block, len(available_blockers), len(available_targets))):
blockers[available_blockers[i]] = available_targets[i]
# Positions: place targets first, then put blockers in front of them
positions = {}
x_slots = [-0.15, 0.0, 0.15, -0.08, 0.08]
random.shuffle(x_slots)
slot_idx = 0
for obj in present:
if obj in blockers.values():
# target that gets blocked — place it further back
positions[obj] = (x_slots[slot_idx % len(x_slots)], -0.05)
else:
positions[obj] = (x_slots[slot_idx % len(x_slots)], 0.05)
slot_idx += 1
# Blocker slightly in front of what it blocks
for blocker, blocked in blockers.items():
tx, ty = positions[blocked]
positions[blocker] = (tx + random.uniform(-0.03, 0.03), ty + 0.08)
# Constraint
constraint = random.choice(CONSTRAINTS)
hidden_traits = {}
for obj in targets_list:
# Keep trait labels simple and interpretable for LLM reasoning.
hidden_traits[obj] = random.choice(["fragile", "heavy", "standard"])
deadlines = {}
if targets_list and random.random() < 0.6:
urgent_obj = random.choice(targets_list)
deadlines[urgent_obj] = random.randint(5, 10)
# Generate instruction
instruction = _build_instruction(target_bins, constraint, hidden_traits, deadlines)
return ScenarioConfig(
objects=present,
targets=target_bins,
blockers=blockers,
distractors=distractors,
constraint=constraint,
instruction=instruction,
positions=positions,
hidden_traits=hidden_traits,
deadlines=deadlines,
)
def _build_instruction(target_bins: dict[str, str], constraint: Optional[str],
hidden_traits: dict[str, str], deadlines: dict[str, int]) -> str:
parts = []
for obj, bin_ in target_bins.items():
display = OBJECT_COLORS.get(obj, obj.replace("_block", ""))
# Use bare display name for non-block objects (professional packs)
label = f"the {display} block" if obj.endswith("_block") else f"the {display}"
parts.append(f"{label} in bin {bin_}")
if len(parts) == 1:
base = f"Place {parts[0]}."
else:
base = "Place " + ", then ".join(parts) + "."
if constraint == "fragile_first":
base += " Handle fragile items first."
elif constraint == "heavy_last":
base += " Move heavy items last."
elif constraint == "urgent_first":
base += " Prioritize urgent items first."
if deadlines:
for obj, step in deadlines.items():
display = OBJECT_COLORS.get(obj, obj.replace("_block", ""))
label = f"the {display} block" if obj.endswith("_block") else f"the {display}"
base += f" Place {label} by step {step}."
if hidden_traits:
base += " Some object traits are hidden until you inspect the scene."
return base
SCENARIO_PACKS = {
"default": OBJECT_NAMES,
# Professional task skins — same mechanics, domain-appropriate names
"warehouse": ["fragile_package", "heavy_pallet", "urgent_parcel", "standard_box", "hazmat_drum"],
"pharmacy": ["morphine_vial", "saline_bag", "insulin_pen", "blood_sample", "contrast_agent"],
"lab": ["reagent_alpha", "catalyst_beta", "sample_gamma", "solvent_delta", "enzyme_epsilon"],
}
# Color/display name for each object in each pack
OBJECT_COLORS.update({
"fragile_package": "fragile package", "heavy_pallet": "heavy pallet",
"urgent_parcel": "urgent parcel", "standard_box": "standard box",
"hazmat_drum": "hazmat drum",
"morphine_vial": "morphine vial", "saline_bag": "saline bag",
"insulin_pen": "insulin pen", "blood_sample": "blood sample",
"contrast_agent": "contrast agent",
"reagent_alpha": "reagent-α", "catalyst_beta": "catalyst-β",
"sample_gamma": "sample-γ", "solvent_delta": "solvent-δ",
"enzyme_epsilon": "enzyme-ε",
})
|