game2player / app.py
multimodalart's picture
multimodalart HF Staff
Upload app.py with huggingface_hub
20b26ed verified
Raw History Blame Contribute Delete
25.7 kB
"""Game2Player demo: game screenshot + task -> next keyboard/mouse action.
Game2Player (https://arxiv.org/abs/2611.24154) is a visual game-playing agent
fine-tuned from Qwen3.5. This Space runs the released 9B checkpoint with
transformers, replicating the authors' vLLM serving recipe (bf16, max_pixels
1048576, thinking off, temperature 1.0) and their PlayerBench prompt/decoder.
"""
import json
import math
import re
import time
from pathlib import Path
import spaces # must be imported before torch / transformers
import torch
import gradio as gr
from PIL import Image, ImageDraw
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration
MODEL_ID = "game2player/Game2Player-9B"
VIEWPORT_W, VIEWPORT_H = 1280, 720
# Authors' vLLM --mm-processor-kwargs max_pixels=1048576 (min_pixels default 3136).
# New-style processor kwarg: size={'longest_edge': max_pixels, 'shortest_edge': min_pixels}.
IMAGE_SIZE = {"longest_edge": 1048576, "shortest_edge": 3136}
EXAMPLES_DIR = Path(__file__).parent / "examples"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
MODEL_ID,
dtype="auto", # bfloat16 per config.json
attn_implementation="sdpa",
).to("cuda")
model.eval()
# ---------------------------------------------------------------------------
# Prompt construction, ported 1:1 from the Game2Player repo
# (agents/qwen_prompting.py + catalog/prompt_templates/computer_use_agent_template.j2)
# ---------------------------------------------------------------------------
MEMORY_NOTE = (
"The user message may contain explicitly labelled previous steps. Treat them "
"as chronological short-term context; act on the labelled current step."
)
CANONICAL_ACTIONS = (
"click", "click_hold", "mouse_move", "drag", "scroll", "type",
"press_key", "press_keys", "set_keys", "key_chunk", "wait",
)
KEY_ALIASES = {
"arrowleft": "ArrowLeft", "left": "ArrowLeft", "arrowright": "ArrowRight",
"right": "ArrowRight", "arrowup": "ArrowUp", "up": "ArrowUp",
"arrowdown": "ArrowDown", "down": "ArrowDown", "space": "Space",
"spacebar": "Space", "enter": "Enter", "return": "Enter", "esc": "Escape",
"escape": "Escape", "tab": "Tab", "backspace": "Backspace",
"delete": "Delete", "del": "Delete", "shift": "Shift",
"shiftleft": "ShiftLeft", "shiftright": "ShiftRight",
"control": "Control", "ctrl": "Control", "controlleft": "ControlLeft",
"controlright": "ControlRight", "alt": "Alt", "altleft": "AltLeft",
"altright": "AltRight", "slash": "/", "period": ".", "comma": ",",
"quote": "'", "apostrophe": "'", "semicolon": ";", "backslash": "\\",
"bracketleft": "[", "bracketright": "]", "minus": "-", "equal": "=",
}
def normalize_key(key):
text = str(key).strip()
if not text:
return None
if len(text) == 1:
return text.lower()
return KEY_ALIASES.get(text.lower(), text)
NOTE_RULE = (
"In `note`, write one short sentence on why you chose this action; your later "
"steps show it."
)
KEYBOARD_RULE = (
"\npress_keys presses its keys simultaneously, then releases them; repeated key "
"names do not create repeated taps. set_keys keeps its keys held until another "
"keyboard action changes them; empty keys releases them. wait advances time "
"without releasing held keys."
)
TC_OPEN = "<tool" + "_call>"
TC_CLOSE = "</tool" + "_call>"
def build_output_format_block(allowed_keys, thinking=False):
"""Port of build_qwen_action_instruction with the repo's ACTION_CONTRACT
(max_actions=1, duration_policy='optional', note=True)."""
keys = sorted({str(k).strip() for k in allowed_keys if str(k).strip()})
lines = []
if keys:
lines.extend([
"- press_key: key, optional duration.",
"- press_keys: keys, optional duration; press simultaneously, then release.",
"- set_keys: keys, optional duration; keep these keys held after the action. Empty keys releases all held keys.",
"- key_chunk: ordered steps containing keys and duration, plus release_after.",
])
lines.extend([
"- type: text.",
"- click: coordinate, optional button (left/right/middle, default left); omit duration.",
"- click_hold: coordinate, optional button and duration.",
"- mouse_move: coordinate.",
"- drag: start_coordinate, end_coordinate, optional button, duration, and steps.",
"- scroll: delta_x and delta_y.",
"- wait: optional duration; advance time without changing held keys.",
"Every action also takes note: one short sentence on why.",
])
example_action = (
{"action": "press_key", "key": keys[0]} if keys
else {"action": "click", "coordinate": [500, 500]}
)
example_action["note"] = "why this action"
example = json.dumps(
{"name": "computer_use", "arguments": example_action},
ensure_ascii=False, separators=(",", ":"),
)
key_rule = (
"Allowed key values: " + ", ".join(json.dumps(k) for k in keys) + ". Do not use any other value as a key."
if keys else "This game has no keyboard controls; do not emit a keyboard action."
)
response_rule = (
"Think in the thinking channel, then emit only the " + TC_OPEN + " as the final "
"answer, with no prose outside the tool-call tags."
if thinking else
"Encode the action plan as one JSON object inside " + TC_OPEN + TC_CLOSE + ". "
"Do not write prose outside the tool-call tags or a Markdown fence."
)
instructions = (
"Return exactly one action.\n"
"Duration is optional, in seconds, at most 10. Without a duration, press_key, "
"press_keys and click are 0.05-second taps, and wait, click_hold and drag "
"last the game's default time."
+ KEYBOARD_RULE
+ "\n" + NOTE_RULE
)
return "# Computer-use output\n" + instructions + "\n" + response_rule + (
" Put every action parameter directly inside\n"
"`arguments`; never put function-call syntax inside the `action` string.\n\n"
"Example:\n" + TC_OPEN + example + TC_CLOSE + "\n\n"
"Coordinates are [x, y] in a normalized 0-1000 range mapped to the "
f"{VIEWPORT_W}x{VIEWPORT_H} viewport.\n"
+ key_rule + "\n"
"Mouse actions are allowed.\n\n"
"Available actions and fields:\n" + chr(10).join(lines)
)
def build_system_prompt(game_rules, role_description, control_notes, task, allowed_keys):
role_block = "\n\n".join(
part.strip() for part in (role_description, control_notes, MEMORY_NOTE) if part and part.strip()
)
return f"""You are an expert game agent. Use the current screenshot and any explicitly
labelled previous steps to choose the next action that best advances the task.
Follow the game rules and control restrictions exactly.
# Game Rules
{game_rules.strip()}
# Role and Controls
{role_block}
# Task Instruction
{task.strip()}
# Output Format
{build_output_format_block(allowed_keys)}"""
# ---------------------------------------------------------------------------
# Output decoding, ported from agents/codecs.py (parse_qwen_action path)
# ---------------------------------------------------------------------------
TOOL_CALL_RE = re.compile(
r"<tool(?:\s+|_)call>\s*(\{[\s\S]*?\})\s*</tool(?:\s+|_)call>", re.IGNORECASE
)
def _point(value, width, height):
"""0-1000 normalized coords -> viewport pixels (agents/codecs._point)."""
if isinstance(value, dict):
x, y = value.get("x"), value.get("y")
elif isinstance(value, (list, tuple)) and len(value) >= 2:
x, y = value[0], value[1]
elif isinstance(value, str):
nums = re.findall(r"-?\d+(?:\.\d+)?", value)
if len(nums) < 2:
return None
x, y = nums[0], nums[1]
else:
return None
try:
x, y = float(x), float(y)
except (TypeError, ValueError):
return None
if not (math.isfinite(x) and math.isfinite(y)):
return None
return (
max(0.0, min(1000.0, x)) / 1000.0 * width,
max(0.0, min(1000.0, y)) / 1000.0 * height,
)
def canonicalize_action(args, width=VIEWPORT_W, height=VIEWPORT_H):
"""Port of _canonicalize_args (normalized=True) for one action dict."""
action_type = str(args.get("action") or "").lower()
if action_type not in CANONICAL_ACTIONS:
return None
if action_type in {"click", "click_hold", "mouse_move"}:
point = _point(
args.get("coordinate") or args.get("point") or args.get("start_box")
or ([args.get("x"), args.get("y")] if "x" in args and "y" in args else None),
width, height,
)
if point is None:
return None
out = {"action": action_type, "x": point[0], "y": point[1]}
if args.get("button") in ("left", "right", "middle"):
out["button"] = args["button"]
return out
if action_type == "drag":
start = _point(args.get("start_coordinate") or args.get("start"), width, height)
end = _point(args.get("end_coordinate") or args.get("end") or args.get("coordinate"), width, height)
if start is None or end is None:
return None
return {"action": "drag", "x1": start[0], "y1": start[1],
"x2": end[0], "y2": end[1]}
if action_type in {"press_key", "press_keys", "set_keys"}:
raw = args.get("keys", args.get("key"))
if isinstance(raw, str):
keys = [k for part in re.split(r"[,+\s]+", raw.strip()) if part for k in [normalize_key(part)] if k]
elif isinstance(raw, (list, tuple)):
keys = [k for k in (normalize_key(i) for i in raw) if k]
else:
keys = []
if not keys:
return None
if action_type == "press_key":
return {"action": "press_key", "key": keys[0]}
return {"action": action_type, "keys": keys}
if action_type == "type":
text = args.get("text", args.get("content"))
if text in (None, ""):
return None
return {"action": "type", "text": str(text)}
if action_type == "scroll":
def _f(v):
try:
n = float(v)
except (TypeError, ValueError):
return 0.0
return n if math.isfinite(n) else 0.0
return {"action": "scroll", "delta_x": _f(args.get("delta_x")), "delta_y": _f(args.get("delta_y"))}
if action_type == "key_chunk":
return None # rare in practice; skip complex chunk parsing in the demo
return {"action": "wait"} if action_type == "wait" else None
def parse_model_output(text):
"""Port of parse_qwen_action: tool-call JSON -> canonical action; fall back to raw JSON."""
for match in TOOL_CALL_RE.finditer(text or ""):
try:
payload = json.loads(match.group(1))
except json.JSONDecodeError:
continue
if not isinstance(payload, dict):
continue
args = payload.get("arguments")
if isinstance(args, str):
try:
args = json.loads(args)
except json.JSONDecodeError:
continue
if not isinstance(args, dict):
continue
nested = args.get("actions")
candidates = nested if isinstance(nested, list) and nested else [args]
for candidate in candidates:
if not isinstance(candidate, dict):
continue
action = canonicalize_action(candidate)
if action is not None:
note = str(
candidate.get("note") or candidate.get("reasoning")
or candidate.get("reason") or ""
).strip()
return action, note
final = (text or "").strip()
if final.startswith("```json"):
final = final[len("```json"):].strip()
if final.endswith("```"):
final = final[:-3].strip()
if final.startswith("{"):
try:
payload = json.loads(final)
except json.JSONDecodeError:
payload = None
if isinstance(payload, dict):
args = payload.get("arguments") or payload.get("action") or payload
if isinstance(args, dict):
action = canonicalize_action(args)
if action is not None:
note = str(
args.get("note") or args.get("reasoning")
or args.get("reason") or ""
).strip()
return action, note
return None, None
# ---------------------------------------------------------------------------
# PlayerBench example presets (games bundled in the Game2Player repo, Apache-2.0)
# ---------------------------------------------------------------------------
# Exact strings from benchmark/playerbench/tasks/{107_simon-sequence,105_memory-match}.yaml
PRESETS = {
"simon": dict(
screenshot=str(EXAMPLES_DIR / "simon_sequence.png"),
task="Repeat Rose Circuit exactly.",
game_rules=(
"You are playing Simon Sequence, a deterministic signal-memory game.\n\n"
"## Objective\n- Repeat the displayed color sequence in exact order.\n\n"
"## Rules\n"
"- The assigned sequence is displayed before input begins and is public to the player.\n"
"- Click one color at a time from left to right.\n"
"- A wrong color fails the attempt; the task ends successfully after the final correct color.\n"
"- The assigned level is selected by the task URL and starts immediately."
),
role_description="You are repeating the displayed signal. Memorize its order and enter it without mistakes.",
control_notes=(
"- Left-click the color buttons.\n"
"- Follow the numbered signal chips from left to right.\n"
"- The assigned level starts immediately; do not select another level."
),
allowed_keys="",
),
"memory": dict(
screenshot=str(EXAMPLES_DIR / "memory_match.png"),
task="Complete Dawn Sprint by finding all three symbol pairs.",
game_rules=(
"You are playing Memory Match, a deterministic offline concentration game. The assigned "
"memory file is already loaded and ready to play.\n\n"
"## Objective\n- Find every matching pair of symbols on the board.\n\n"
"## Rules\n"
"- Click a face-down card to reveal it, then click a second face-down card.\n"
"- Matching symbols stay face up and count as a pair.\n"
"- A mismatch remains visible until you click either revealed card to hide both, then you may continue.\n"
"- The board is deterministic for each URL level and contains no remote assets.\n"
"- Press R or click Restart to reset the assigned memory file."
),
role_description=(
"You are the memory solver. Reveal cards, remember every visible symbol and its "
"position, then recover all pairs with as few mismatches as possible."
),
control_notes=(
"- Click the center of any face-down card to reveal it.\n"
"- After a mismatch, click either of the two revealed cards to hide both before selecting another pair.\n"
"- A matched card stays face up and does not need to be clicked again.\n"
"- The assigned level is already open from the URL; do not select another level.\n"
"- R restarts the assigned memory file."
),
allowed_keys="r",
),
}
# ---------------------------------------------------------------------------
# Inference
# ---------------------------------------------------------------------------
def annotate(image, action):
"""Draw the predicted click/drag point on a copy of the screenshot."""
canvas = image.copy().convert("RGB")
draw = ImageDraw.Draw(canvas)
if action.get("action") in {"click", "click_hold", "mouse_move"}:
x, y = action.get("x"), action.get("y")
if x is not None and y is not None:
r = 12
draw.ellipse([x - r, y - r, x + r, y + r], outline=(255, 60, 60), width=4)
draw.line([x - r - 6, y, x + r + 6, y], fill=(255, 60, 60), width=2)
draw.line([x, y - r - 6, x, y + r + 6], fill=(255, 60, 60), width=2)
elif action.get("action") == "drag":
x1, y1, x2, y2 = action.get("x1"), action.get("y1"), action.get("x2"), action.get("y2")
if None not in (x1, y1, x2, y2):
draw.line([x1, y1, x2, y2], fill=(255, 60, 60), width=4)
draw.ellipse([x1 - 10, y1 - 10, x1 + 10, y1 + 10], outline=(255, 60, 60), width=3)
draw.ellipse([x2 - 10, y2 - 10, x2 + 10, y2 + 10], outline=(255, 200, 0), width=3)
return canvas
def describe_action(action, note):
"""One-line human rendering of the canonical action."""
if action is None:
return None
a = action["action"]
if a in {"click", "click_hold", "mouse_move"}:
text = f"{a} at pixel ({int(action['x'])}, {int(action['y'])})"
if a == "click":
text += f" [normalized 0-1000: ({round(action['x'] / VIEWPORT_W * 1000)}, {round(action['y'] / VIEWPORT_H * 1000)})]"
elif a == "drag":
text = f"drag from ({int(action['x1'])}, {int(action['y1'])}) to ({int(action['x2'])}, {int(action['y2'])})"
elif a == "press_key":
text = f"press_key {action['key']}"
elif a in {"press_keys", "set_keys"}:
text = f"{a} {'+'.join(action['keys'])}"
elif a == "type":
text = f'type "{action["text"]}"'
elif a == "scroll":
text = f"scroll dx={action['delta_x']} dy={action['delta_y']}"
else:
text = "wait"
if note:
text += f" — {note}"
return text
def _estimate_duration(screenshot, *args, **kwargs):
"""ZeroGPU duration: measured 3.6s at 512 tokens non-thinking; thinking needs more."""
thinking = args[4] if len(args) > 4 else False
max_new_tokens = args[5] if len(args) > 5 else 512
if thinking:
return 90
return min(90, 20 + int(max_new_tokens) * 0.05)
@spaces.GPU(duration=_estimate_duration)
def predict_next_action(
screenshot,
task,
game_rules=PRESETS["simon"]["game_rules"],
role_description=PRESETS["simon"]["role_description"],
control_notes=PRESETS["simon"]["control_notes"],
allowed_keys="",
thinking=False,
max_new_tokens=512,
temperature=1.0,
seed=-1,
progress=gr.Progress(track_tqdm=True),
):
"""Predict the next game action from a screenshot and task description.
Args:
screenshot: current game screenshot (any size; rescaled to ~1M pixels as in training).
task: the task instruction, e.g. "Repeat Warm Signal exactly."
game_rules: the game's rules block (pre-filled from PlayerBench examples).
role_description: the player role description.
control_notes: notes on how controls work.
allowed_keys: comma-separated list of keys the game accepts (empty for mouse-only).
thinking: let the model reason in a thinking channel before the tool call.
max_new_tokens: generation budget.
temperature: sampling temperature.
seed: random seed; -1 keeps sampling random.
Returns:
(annotated screenshot, action summary, raw model output, elapsed seconds).
"""
if screenshot is None:
raise gr.Error("Please provide a game screenshot first.")
# API callers may omit the preset fields (None) — fall back to the Simon preset.
if not game_rules:
game_rules = PRESETS["simon"]["game_rules"]
if role_description is None:
role_description = PRESETS["simon"]["role_description"]
if control_notes is None:
control_notes = PRESETS["simon"]["control_notes"]
if task is None:
task = PRESETS["simon"]["task"]
if isinstance(screenshot, str):
image = Image.open(screenshot).convert("RGB")
else:
image = screenshot.convert("RGB")
keys = [k.strip() for k in re.split(r"[,+\s]+", (allowed_keys or "").strip()) if k.strip()]
system_prompt = build_system_prompt(game_rules, role_description, control_notes, task, keys)
messages = [
{"role": "system", "content": [{"type": "text", "text": system_prompt}]},
{"role": "user", "content": [
{"type": "text", "text": "# Current step\nCurrent game screen:\n"},
{"type": "image", "image": image},
{"type": "text", "text": "\nReturn the next action."},
]},
]
torch.manual_seed(seed if seed is not None and seed >= 0 else int(time.time()))
inputs = processor.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_dict=True,
return_tensors="pt",
enable_thinking=bool(thinking),
processor_kwargs={"size": IMAGE_SIZE},
).to("cuda")
start = time.perf_counter()
with torch.inference_mode():
generated = model.generate(
**inputs,
max_new_tokens=int(max_new_tokens),
temperature=float(temperature),
do_sample=float(temperature) > 0,
top_p=1.0,
)
elapsed = time.perf_counter() - start
trimmed = generated[0][inputs["input_ids"].shape[1]:]
raw = processor.decode(trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)
action, note = parse_model_output(raw)
summary = describe_action(action, note)
if summary is None:
summary = "Could not parse an action from the model output — see the raw output below."
annotated = annotate(image, action) if action and action.get("action") in {
"click", "click_hold", "mouse_move", "drag",
} else image
return annotated, summary, raw, f"{elapsed:.1f}s"
with gr.Blocks(title="Game2Player — Visual Game-Playing Agent") as demo:
gr.Markdown(
f"""
# 🎮 Game2Player — Visual Game-Playing Agent
**Game2Player: Scaling Gameplay Experience for Visual Agents through Environment-Player Co-Development**
([paper](https://arxiv.org/abs/2611.24154) · [code](https://github.com/game2player/Game2Player) · [model card](https://huggingface.co/{MODEL_ID}))
This demo runs the released **Game2Player-9B** checkpoint (Qwen3.5-9B fine-tune) with transformers,
reproducing the authors' serving recipe (bf16, ≤1M pixels per screenshot, temperature 1.0).
Give it a **game screenshot** and the **task + rules**, and it predicts the next
keyboard or mouse action as a `computer_use` tool call.
The two presets are real tasks from **PlayerBench**, the paper's benchmark — screenshots
captured from the benchmark's bundled games (`107_simon-sequence`, `105_memory-match`, Apache-2.0).
"""
)
with gr.Row():
with gr.Column():
screenshot = gr.Image(
type="pil", label="Game screenshot", height=420,
sources=["upload", "clipboard"],
)
task = gr.Textbox(
label="Task instruction", lines=1,
value="Repeat Rose Circuit exactly.",
)
with gr.Accordion("Game rules and controls (advanced)", open=False):
game_rules = gr.Textbox(label="Game rules", lines=8, value=PRESETS["simon"]["game_rules"])
role_description = gr.Textbox(label="Role description", lines=2, value=PRESETS["simon"]["role_description"])
control_notes = gr.Textbox(label="Control notes", lines=3, value=PRESETS["simon"]["control_notes"])
allowed_keys = gr.Textbox(label="Allowed keys (comma-separated; empty = mouse-only)", value="")
with gr.Accordion("Generation settings", open=False):
thinking = gr.Checkbox(label="Enable thinking channel", value=False)
max_new_tokens = gr.Slider(64, 2048, value=512, step=64, label="Max new tokens")
temperature = gr.Slider(0.0, 2.0, value=1.0, step=0.1, label="Temperature")
seed = gr.Number(value=-1, label="Seed (-1 = random)")
predict_btn = gr.Button("Predict next action", variant="primary")
with gr.Column():
annotated_out = gr.Image(type="pil", label="Predicted action on screen", height=420)
action_out = gr.Textbox(label="Next action", lines=2)
raw_out = gr.Textbox(label="Raw model output", lines=6)
elapsed_out = gr.Textbox(label="Inference time", show_label=False)
predict_btn.click(
predict_next_action,
inputs=[screenshot, task, game_rules, role_description, control_notes,
allowed_keys, thinking, max_new_tokens, temperature, seed],
outputs=[annotated_out, action_out, raw_out, elapsed_out],
api_name="predict_next_action",
)
example_inputs = [screenshot, task, game_rules, role_description, control_notes, allowed_keys]
gr.Examples(
examples=[
[p["screenshot"], p["task"], p["game_rules"], p["role_description"],
p["control_notes"], p["allowed_keys"]]
for p in (PRESETS["simon"], PRESETS["memory"])
],
inputs=example_inputs,
fn=predict_next_action,
outputs=[annotated_out, action_out, raw_out, elapsed_out],
cache_examples=True,
cache_mode="lazy",
label="PlayerBench examples (real tasks from the paper's benchmark)",
)
demo.launch(mcp_server=True, theme=gr.themes.Citrus())