| """ |
| VLM evaluation of all 10 conflict experiments using Gemini-2.5-Flash. |
| Reads existing videos and success_rate.txt; NEVER modifies existing files. |
| |
| Adapted for this environment: |
| - model arg: gr00t -> /workspace/groot_eval/results |
| genie -> /workspace/groot_eval/results_genie |
| - layout: <root>/<exp>/experiments/ood_<idx>_<exp>_<pi>_<pj>_<rt>_<ts>/ |
| - output: /workspace/groot_eval/results/vlm_eval/<model>/vlm_eval_<exp>.jsonl |
| - resumable: skips runs already present in the output jsonl |
| Usage: python vlm_eval.py <gr00t|genie> [exp ...] [--limit N] |
| """ |
|
|
| import base64, json, re, sys, time, threading |
| from concurrent.futures import ThreadPoolExecutor |
| from pathlib import Path |
| import requests |
|
|
| WORKERS = 10 |
|
|
| API_KEY = "sk-12YA7oNA-9gl9qn2oedejw" |
| API_URL = "https://inference-api.nvidia.com/v1/chat/completions" |
| MODEL = "gcp/google/gemini-2.5-flash" |
|
|
| ROOTS = { |
| "gr00t": Path("/workspace/groot_eval/results"), |
| "genie": Path("/workspace/groot_eval/results_genie"), |
| } |
| OUT_BASE = Path("/workspace/groot_eval/results/vlm_eval") |
|
|
| VERBS = ["lift", "grasp", "push", "pull", "rotate", "slide"] |
| COLORS = ["red", "yellow", "blue", "orange", "green", "black"] |
| SHAPES = ["cube", "sphere", "cup", "car", "pyramid", "star"] |
| SIZES = ["smallest", "second smallest", "middle-sized", "second largest", "largest", "biggest"] |
| SPATIALS= ["on the left", "on the right", "in the middle", "in front", "at the back", "in the center"] |
|
|
| ALL_EXPERIMENTS = [ |
| "verb_color", "verb_object", "verb_size", "verb_spatial", |
| "color_object", "size_object", "color_size", "color_spatial", |
| "spatial_size", "spatial_object", |
| ] |
|
|
| |
|
|
| def prompt_verb_experiment(instruction, pair_i, pair_j, factor2): |
| verb_i = VERBS[pair_i]; verb_j = VERBS[pair_j] |
| if factor2 == "color": |
| factor2_desc = f"color '{COLORS[pair_j]}' (which the robot was trained to interact with using '{verb_j}')" |
| elif factor2 == "shape": |
| factor2_desc = f"shape '{SHAPES[pair_j]}' (which the robot was trained to interact with using '{verb_j}')" |
| elif factor2 == "size": |
| factor2_desc = f"size '{SIZES[pair_j]}' (which the robot was trained to interact with using '{verb_j}')" |
| elif factor2 == "spatial": |
| factor2_desc = f"spatial position '{SPATIALS[pair_j]}' (which the robot was trained to interact with using '{verb_j}')" |
| else: |
| factor2_desc = f"factor {pair_j} (trained verb: '{verb_j}')" |
| return f"""Watch this robot manipulation video carefully. |
| |
| **Instruction given to the robot:** "{instruction}" |
| |
| This is a CONFLICT experiment about verb bias. The robot was trained with specific verb-attribute pairings: |
| - Verb "{verb_i}" is paired with one attribute set (instruction verb) |
| - Verb "{verb_j}" is paired with {factor2_desc} |
| |
| The robot must choose which action to perform. Observe carefully. |
| |
| Answer in this EXACT format: |
| ROBOT_ACTION: [describe what physical action the robot arm performed in 1 sentence] |
| ACTION_TYPE: ["{verb_i}" if robot performed the instructed action, "{verb_j}" if robot performed the competing action, "other" if neither] |
| FACTOR_FOLLOWED: ["verb" if robot followed the instructed verb '{verb_i}', "{factor2}" if robot followed the {factor2} factor (doing '{verb_j}'), "neither" if unclear] |
| CONFIDENCE: [high / medium / low] |
| REASONING: [1-2 sentences explaining your judgment]""" |
|
|
|
|
| def prompt_two_object_experiment(instruction, pair_i, pair_j, factor1, factor2): |
| def describe_factor(f, k): |
| if f == "color": return f"color='{COLORS[k]}'" |
| if f == "shape": return f"shape='{SHAPES[k]}'" |
| if f == "size": return f"size='{SIZES[k]}'" |
| if f == "spatial": return f"position='{SPATIALS[k]}'" |
| return f"{f}={k}" |
| obj_a_desc = describe_factor(factor1, pair_i) |
| obj_b_desc = describe_factor(factor2, pair_j) |
| return f"""Watch this robot manipulation video carefully. |
| |
| **Instruction given to the robot:** "{instruction}" |
| |
| This is a CONFLICT experiment about factor bias. The scene contains two objects: |
| - Object A: has {obj_a_desc} — matches the '{factor1}' factor in the instruction |
| - Object B: has {obj_b_desc} — matches the '{factor2}' factor in the instruction |
| |
| No object has BOTH attributes simultaneously. The robot must choose one. |
| |
| Answer in this EXACT format: |
| ROBOT_ACTION: [describe what the robot arm did in 1 sentence] |
| OBJECT_TOUCHED: [A (matches {factor1}), B (matches {factor2}), or neither] |
| FACTOR_FOLLOWED: ["{factor1}" if robot went for Object A, "{factor2}" if robot went for Object B, "neither"] |
| CONFIDENCE: [high / medium / low] |
| REASONING: [1-2 sentences explaining your judgment]""" |
|
|
|
|
| VERB_EXPS = {"verb_color": "color", "verb_object": "shape", |
| "verb_size": "size", "verb_spatial": "spatial"} |
| TWO_OBJ_EXPS = { |
| "color_object": ("color", "shape"), "size_object": ("size", "shape"), |
| "color_size": ("color", "size"), "color_spatial": ("color", "spatial"), |
| "spatial_size": ("spatial", "size"), "spatial_object": ("spatial", "shape"), |
| } |
|
|
| def get_prompt(experiment, instruction, pair_i, pair_j, run_type): |
| if experiment in VERB_EXPS: |
| return prompt_verb_experiment(instruction, pair_i, pair_j, VERB_EXPS[experiment]) |
| if experiment in TWO_OBJ_EXPS: |
| f1, f2 = TWO_OBJ_EXPS[experiment] |
| return prompt_two_object_experiment(instruction, pair_i, pair_j, f1, f2) |
| return f'Watch this robot video. Instruction: "{instruction}". What did the robot do?' |
|
|
| |
|
|
| def call_vlm(video_path, prompt): |
| with open(video_path, "rb") as f: |
| video_b64 = base64.b64encode(f.read()).decode() |
| payload = {"model": MODEL, "max_tokens": 2000, |
| "messages": [{"role": "user", "content": [ |
| {"type": "text", "text": prompt}, |
| {"type": "image_url", "image_url": {"url": f"data:video/mp4;base64,{video_b64}"}}]}]} |
| for attempt in range(4): |
| try: |
| resp = requests.post(API_URL, |
| headers={"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json"}, |
| json=payload, timeout=120) |
| if resp.status_code == 200: |
| content = resp.json()["choices"][0]["message"]["content"] or "" |
| return parse_response(content) |
| elif resp.status_code == 429: |
| print(" rate-limited 30s", flush=True); time.sleep(30) |
| else: |
| print(f" API {resp.status_code}: {resp.text[:120]}", flush=True); time.sleep(5) |
| except Exception as e: |
| print(f" exc: {e}", flush=True); time.sleep(5) |
| return {"raw": "FAILED", "factor_followed": "error"} |
|
|
|
|
| def parse_response(text): |
| result = {"raw": text} |
| for field in ["ROBOT_ACTION", "OBJECT_TOUCHED", "ACTION_TYPE", |
| "FACTOR_FOLLOWED", "CONFIDENCE", "REASONING"]: |
| m = re.search(rf"{field}:\s*(.+?)(?=\n[A-Z_]+:|$)", text, re.DOTALL) |
| result[field.lower()] = m.group(1).strip() if m else "" |
| return result |
|
|
| |
|
|
| def load_runs(root: Path, experiment: str): |
| expdir = root / experiment / "experiments" |
| best = {} |
| for d in sorted(expdir.glob(f"ood_*_{experiment}_*")): |
| if not d.is_dir(): |
| continue |
| parts = d.name.split("_") |
| try: |
| exp_len = len(experiment.split("_")) |
| idx = int(parts[1]) |
| pair_i = int(parts[2 + exp_len]) |
| pair_j = int(parts[3 + exp_len]) |
| run_type = parts[4 + exp_len] |
| run_name = "_".join(parts[:5 + exp_len]) |
| except (ValueError, IndexError): |
| continue |
| sr = d / "success_rate.txt" |
| existing = "?/1" |
| if sr.exists(): |
| m = re.search(r"_success=(\d+/\d+)", sr.read_text()) |
| if m: existing = m.group(1) |
| best[idx] = {"idx": idx, "pair_i": pair_i, "pair_j": pair_j, |
| "run_type": run_type, "existing_success": existing, |
| "run_name": run_name, "dir": str(d)} |
| return [best[k] for k in sorted(best)] |
|
|
| |
|
|
| def evaluate_experiment(model: str, experiment: str, limit: int = 0): |
| root = ROOTS[model] |
| out_dir = OUT_BASE / model |
| out_dir.mkdir(parents=True, exist_ok=True) |
| out_file = out_dir / f"vlm_eval_{experiment}.jsonl" |
| print(f"\n{'='*60}\n[{model}] Evaluating: {experiment}\nOutput: {out_file}", flush=True) |
|
|
| runs = load_runs(root, experiment) |
| print(f"found {len(runs)} runs", flush=True) |
|
|
| done_names = set() |
| if out_file.exists(): |
| for line in out_file.read_text().splitlines(): |
| try: |
| done_names.add(json.loads(line)["run_name"]) |
| except Exception: |
| pass |
| if done_names: |
| print(f"resume: {len(done_names)} already done, skipping them", flush=True) |
|
|
| factor_counts = {} |
| |
| if out_file.exists(): |
| for line in out_file.read_text().splitlines(): |
| try: |
| f = json.loads(line).get("vlm_factor_followed", "error") |
| factor_counts[f] = factor_counts.get(f, 0) + 1 |
| except Exception: |
| pass |
|
|
| |
| pending = [] |
| for run in runs: |
| if run["run_name"] in done_names: |
| continue |
| run_dir = Path(run["dir"]) |
| video_path = run_dir / "video" / "ep000.mp4" |
| if not run_dir.exists() or not video_path.exists(): |
| print(f" [{run['idx']}] MISSING dir/video", flush=True); continue |
| instruction = "unknown" |
| sr_file = run_dir / "success_rate.txt" |
| if sr_file.exists(): |
| m = re.search(r"instruction='(.+?)'", sr_file.read_text()) |
| if m: instruction = m.group(1) |
| run["_video"] = str(video_path); run["_instruction"] = instruction |
| pending.append(run) |
| if limit and len(pending) >= limit: |
| break |
| print(f"pending={len(pending)} workers={WORKERS}", flush=True) |
|
|
| out_f = open(out_file, "a") |
| lock = threading.Lock() |
| state = {"n": 0} |
|
|
| def work(run): |
| prompt = get_prompt(experiment, run["_instruction"], run["pair_i"], |
| run["pair_j"], run["run_type"]) |
| vlm = call_vlm(Path(run["_video"]), prompt) |
| factor = vlm.get("factor_followed", "error").lower().strip('"\'').strip() |
| record = {"model": model, "experiment": experiment, "run_name": run["run_name"], |
| "idx": run["idx"], "pair_i": run["pair_i"], "pair_j": run["pair_j"], |
| "run_type": run["run_type"], "instruction": run["_instruction"], |
| "existing_success": run["existing_success"], |
| "vlm_factor_followed": factor, |
| "vlm_object_touched": vlm.get("object_touched", ""), |
| "vlm_action_type": vlm.get("action_type", ""), |
| "vlm_robot_action": vlm.get("robot_action", ""), |
| "vlm_confidence": vlm.get("confidence", ""), |
| "vlm_reasoning": vlm.get("reasoning", ""), |
| "vlm_raw": vlm.get("raw", "")} |
| with lock: |
| out_f.write(json.dumps(record) + "\n"); out_f.flush() |
| factor_counts[factor] = factor_counts.get(factor, 0) + 1 |
| state["n"] += 1 |
| print(f" [{model}/{experiment} {state['n']}/{len(pending)}] " |
| f"idx{run['idx']} {factor.upper()} [{vlm.get('confidence','')}]", flush=True) |
|
|
| if pending: |
| with ThreadPoolExecutor(max_workers=WORKERS) as ex: |
| list(ex.map(work, pending)) |
| out_f.close() |
|
|
| total = sum(factor_counts.values()) |
| with open(out_dir / f"vlm_eval_{experiment}_summary.txt", "w") as f: |
| f.write(f"# VLM Eval: {model} {experiment}\nmodel: {MODEL}\ntotal_evaluated: {total}\n\n") |
| for fac, cnt in sorted(factor_counts.items(), key=lambda x: -x[1]): |
| f.write(f"{fac}: {cnt}/{total} ({100*cnt/max(1,total):.1f}%)\n") |
| print(f" [{model}/{experiment}] +{state['n']} new, totals={factor_counts}", flush=True) |
| return factor_counts |
|
|
|
|
| if __name__ == "__main__": |
| args = [a for a in sys.argv[1:] if not a.startswith("--")] |
| limit = 0 |
| for i, a in enumerate(sys.argv[1:], start=1): |
| if a.startswith("--limit"): |
| limit = int(a.split("=")[1]) if "=" in a else int(sys.argv[i+1]) |
| if a.startswith("--workers"): |
| WORKERS = int(a.split("=")[1]) if "=" in a else int(sys.argv[i+1]) |
| model = args[0] if args else "gr00t" |
| exps = [a for a in args[1:] if a in ALL_EXPERIMENTS] or ALL_EXPERIMENTS |
| assert model in ROOTS, f"model must be gr00t|genie, got {model}" |
| results = {} |
| for e in exps: |
| results[e] = evaluate_experiment(model, e, limit=limit) |
| print("\n" + "="*60 + f"\n[{model}] ALL DONE") |
| for e, c in results.items(): |
| t = sum(c.values()); top = max(c, key=c.get) if c else "NA" |
| print(f" {e}: dominant={top} ({100*c.get(top,0)/max(1,t):.1f}%)") |
|
|