Nabidnur's picture
RoboRACE: code, cached agent outputs, results, media, SO-101 renders (part 5)
99e41a8 verified
Raw History Blame Contribute Delete
3.72 kB
"""Small curated benchmark for the task planner.
Each case: scene objects, instruction, and a gold list of (skill, object, location-keyword)
steps; `None` gold means the planner must declare the task infeasible.
Matching is order-sensitive on (skill, object); location only needs to contain the keyword.
"""
from __future__ import annotations
CASES: list[dict] = [
# --- single-step ---
{"scene": ["red cube", "blue bowl"], "instr": "pick up the red cube",
"gold": [("pick", "red cube", None)]},
{"scene": ["drawer", "green cube"], "instr": "open the drawer",
"gold": [("open", "drawer", None)]},
# --- pick & place ---
{"scene": ["red cube", "blue bowl", "green cube"], "instr": "put the red cube in the blue bowl",
"gold": [("pick", "red cube", None), ("place", "red cube", "bowl")]},
{"scene": ["yellow cube", "green cube", "plate"], "instr": "stack the yellow cube on the green cube",
"gold": [("pick", "yellow cube", None), ("stack", "yellow cube", "green")]},
# --- multi-object / long-horizon ---
{"scene": ["red cube", "green cube", "blue bowl"], "instr": "put both cubes in the bowl, red first",
"gold": [("pick", "red cube", None), ("place", "red cube", "bowl"),
("pick", "green cube", None), ("place", "green cube", "bowl")]},
{"scene": ["drawer", "red cube"], "instr": "put the red cube in the drawer and close it",
"gold": [("open", "drawer", None), ("pick", "red cube", None),
("place", "red cube", "drawer"), ("close", "drawer", None)]},
{"scene": ["red cube", "green cube", "blue cube"],
"instr": "build a tower: blue at the bottom, then green, then red on top",
"gold": [("pick", "green cube", None), ("stack", "green cube", "blue"),
("pick", "red cube", None), ("stack", "red cube", "green")]},
# --- implicit / commonsense ---
{"scene": ["apple", "banana", "fruit bowl", "screwdriver"], "instr": "tidy up the fruit",
"gold": [("pick", "apple", None), ("place", "apple", "bowl"),
("pick", "banana", None), ("place", "banana", "bowl")], "any_order": True},
{"scene": ["mug", "sponge", "sink"], "instr": "I'm done with my mug, put it where dirty dishes go",
"gold": [("pick", "mug", None), ("place", "mug", "sink")]},
{"scene": ["red cube", "blue bowl"], "instr": "when you're finished, return the arm to its rest pose",
"gold": [("home", None, None)]},
# --- must refuse (infeasible) ---
{"scene": ["red cube", "blue bowl"], "instr": "put the green cube in the bowl", "gold": None},
{"scene": ["red cube", "blue bowl"], "instr": "fly to the kitchen and make coffee", "gold": None},
]
def _norm(s: str | None) -> str:
return (s or "").lower().strip()
def step_match(pred, gold) -> bool:
skill, obj, loc = gold
if pred.skill != skill:
return False
if obj and _norm(pred.object) != obj:
return False
return not loc or loc in _norm(pred.location)
def score(plan, gold, any_order: bool = False) -> bool:
if gold is None:
return len(plan.steps) == 0
# tolerate an optional trailing "home" the planner may add
steps = [s for s in plan.steps if s.skill != "home"] if not any(g[0] == "home" for g in gold) else plan.steps
if len(steps) != len(gold):
return False
if any_order: # pick/place pairs may come in any object order
pairs_p = sorted(zip(steps[::2], steps[1::2]), key=lambda p: _norm(p[0].object))
pairs_g = sorted(zip(gold[::2], gold[1::2]), key=lambda p: p[0][1])
return all(step_match(a, b) and step_match(c, d) for (a, c), (b, d) in zip(pairs_p, pairs_g))
return all(step_match(p, g) for p, g in zip(steps, gold))