#!/usr/bin/env python3 """Independent, offline checks for the Franka/FT OXT export. This verifies preservation and declared numeric conversions, not physical sensor calibration, canonical gripper alignment, image content, or source-file hashes. It does not import the exporter or open an SVO/live camera. RGB chunks are sampled at each episode's first/last row unless --scan-rgb-all is requested. """ from __future__ import annotations import argparse from datetime import datetime, timezone import json from pathlib import Path import sys from typing import Any import h5py import numpy as np from scipy.spatial.transform import Rotation TASKS = ("wipe_board", "push_button", "insert_pen", "hang_tape") CAMERAS = {"camera_main": "36087771", "right_wrist_camera": "16478870"} STANDARD_SHAPES = { "timestamps": (), "timestamp_unix_ns": (), "source_control_row": (), "right_arm_joints": (7,), "right_hand_joints": (1,), "right_hand_joints_idx": (1,), "sub_task_instruction": (), "action_joint_position": (7,), "action_eef_pose": (6,), "action_eef_pose_with_ft_offset": (6,), "state_eef_pose_with_ft_offset": (6,), "command_dispatched": (), "right_tactile_data_wrist_torque": (1, 3), "right_tactile_area_wrist_torque": (1,), "right_tactile_sensor_wrist_torque": (), "right_tactile_type_wrist_torque": (), "ft_wrench": (6,), "supplementary_tactile": (6,), } POSE_SOURCES = { "state_eef_pose_with_ft_offset": "observation/robot_state/cartesian_position_with_ft_offset", "action_eef_pose": "action/cartesian_position", "action_eef_pose_with_ft_offset": "action/cartesian_position_with_ft_offset", } def text_value(value: Any) -> str: return value.decode("utf-8") if isinstance(value, bytes) else str(value) def episode_name_matches(original_name: str, exported_name: str) -> bool: """Allow only the exact native name or its documented ':' -> '-' variant.""" return exported_name == original_name or exported_name == original_name.replace(":", "-") def native_key(path: str) -> str: return "native__" + path.replace("/", "__") def strings(values: Any) -> np.ndarray: arr = np.asarray(values) return np.asarray([text_value(v) for v in arr.reshape(-1)]).reshape(arr.shape) def exactly_equal(actual: np.ndarray, expected: np.ndarray) -> bool: if actual.shape != expected.shape: return False if expected.dtype.kind in "OSU": return bool(np.array_equal(strings(actual), strings(expected))) return bool(np.array_equal(actual, expected, equal_nan=True)) class Checks: def __init__(self) -> None: self.errors: list[str] = [] self.count = 0 def require(self, condition: Any, message: str) -> bool: self.count += 1 ok = bool(condition) if not ok: self.errors.append(message) return ok def equal(self, actual: Any, expected: Any, label: str) -> None: self.require(exactly_equal(np.asarray(actual), np.asarray(expected)), label + ": value mismatch") def pose_check(actual: np.ndarray, source: np.ndarray, checks: Checks, label: str) -> None: if not checks.require(actual.shape == source.shape and actual.shape[-1:] == (6,), label + ": pose shape mismatch"): return if not checks.require(np.isfinite(actual).all() and np.isfinite(source).all(), label + ": nonfinite pose"): return checks.require(np.allclose(actual[:, :3], source[:, :3], atol=1e-10, rtol=0), label + ": position changed") expected_rotation = Rotation.from_euler("xyz", source[:, 3:]).as_matrix() exported_rotation = Rotation.from_rotvec(actual[:, 3:]).as_matrix() checks.require(np.allclose(exported_rotation, expected_rotation, atol=1e-10, rtol=0), label + ": Euler-to-rotvec rotation mismatch") roundtrip = Rotation.from_rotvec(Rotation.from_matrix(exported_rotation).as_rotvec()).as_matrix() checks.require(np.allclose(roundtrip, expected_rotation, atol=1e-10, rtol=0), label + ": rotation roundtrip mismatch") def gripper_check(actual: np.ndarray, source_closure: np.ndarray, checks: Checks) -> None: """Independently verify slot 28 opening against untouched native closure.""" checks.equal(actual, (1 - source_closure).reshape(-1, 1), "normalized gripper opening must equal 1 - native closure") checks.require(np.isfinite(actual).all() and np.all((actual >= 0) & (actual <= 1)), "normalized gripper opening must be finite and within [0, 1]") def offset_check(original: np.ndarray, derived: np.ndarray, extension: float, checks: Checks, label: str) -> None: expected = original[:, :3] + Rotation.from_euler("xyz", original[:, 3:]).apply([0.0, 0.0, extension]) checks.require(np.allclose(derived[:, :3], expected, atol=1e-10, rtol=0), label + ": local +Z offset formula mismatch") checks.equal(derived[:, 3:], original[:, 3:], label + ": offset must retain original orientation") def scan_rgb_rows(rgb: Any, span: slice, checks: Checks, label: str) -> int: """Read every RGB row in batches, independently of Zarr's chunk length.""" tested = 0 for pos in range(span.start, span.stop, 32): stop = min(pos + 32, span.stop) image_rows = np.asarray(rgb[pos:stop]) checks.require(image_rows.dtype == np.uint8 and image_rows.shape == (stop - pos, 360, 640, 3), label + ": decoded RGB batch malformed") tested += len(image_rows) return tested def find_audit(root: Path, explicit: Path | None) -> Path | None: if explicit is not None: return explicit for path in (root / "metadata/selected_sources_audit.json", root / "provenance/selected_sources_audit.json", root / "selected_sources_audit.json", root.parent / "oxt_export_work/selected_sources_audit.json"): if path.is_file(): return path return None def validate_episode(data: Any, span: slice, package_root: Path, relative_native: str, episode_id: str, episode_name: str, audit: dict | None, scan_rgb_all: bool) -> dict: checks = Checks() result: dict[str, Any] = {"episode_id": episode_id, "episode_name": episode_name, "native_path": relative_native, "errors": checks.errors} rel = Path(relative_native) if not checks.require(not rel.is_absolute() and ".." not in rel.parts and rel.parts[:1] == ("native",), "native path must be a relative path under native/"): return result source_dir = package_root / rel with h5py.File(source_dir / "trajectory.h5", "r") as source, h5py.File(source_dir / "ft_sensor.h5", "r") as ft: rows = np.asarray(data["source_control_row"][span]) raw_T = int(source["action/joint_position"].shape[0]) result.update(control_rows_exported=int(len(rows)), control_rows_native=raw_T, control_rows_omitted=raw_T - len(rows)) if not checks.require(rows.ndim == 1 and rows.dtype.kind in "iu" and rows.size > 0 and np.all(rows >= 0) and np.all(rows < raw_T) and np.all(np.diff(rows.astype(np.int64)) > 0), "source_control_row must be increasing valid integer indices"): return result rows = rows.astype(np.int64) checks.require(bool(source.attrs.get("complete", False)), "native trajectory incomplete") checks.require(bool(source.attrs.get("success", False)), "native episode is not operator success") checks.require(not text_value(source.attrs.get("collection_error", "")), "native trajectory collection_error is nonempty") checks.require(text_value(source.attrs.get("episode_id", "")) == episode_id, "native episode_id mismatch") original_name = text_value(source.attrs.get("time", "")) result["original_episode_name"] = original_name result["episode_name_normalization"] = ( "unchanged" if episode_name == original_name else "colon_to_hyphen_portable_variant" if episode_name_matches(original_name, episode_name) else "invalid" ) checks.require(episode_name_matches(original_name, episode_name), "native episode_name mismatch: expected exact original or colon-to-hyphen portable variant") checks.require(bool(ft.attrs.get("complete", False)), "native FT stream incomplete") # Check every raw row dataset, including controller diagnostics, actions, # camera calibration and clocks. Strings change storage dtype only. row_paths: list[str] = [] source.visititems(lambda name, obj: row_paths.append(name) if isinstance(obj, h5py.Dataset) and obj.ndim > 0 and obj.shape[0] == raw_T else None) result["native_row_datasets"] = len(row_paths) for path in row_paths: key = native_key(path) if not checks.require(key in data, "missing native field " + key): continue expected = source[path][rows] actual = np.asarray(data[key][span]) checks.equal(actual, expected, key) if expected.dtype.kind not in "OSU": checks.require(actual.dtype == expected.dtype, key + ": numeric dtype changed") if actual.dtype.kind in "fc": checks.require(np.isfinite(actual).all(), key + ": contains nonfinite numeric values") exported_native = {key for key in data.array_keys() if key.startswith("native__")} checks.require(exported_native == {native_key(path) for path in row_paths}, "native field set differs from source row dataset set") def raw(path: str) -> np.ndarray: return np.asarray(source[path][rows]) def out(key: str) -> np.ndarray: return np.asarray(data[key][span]) checks.equal(out("right_arm_joints"), raw("observation/robot_state/joint_positions"), "standard arm state") gripper_check(out("right_hand_joints"), raw("observation/robot_state/gripper_position"), checks) checks.equal(out("action_joint_position"), raw("action/joint_position"), "actual joint targets") for key, path in POSE_SOURCES.items(): pose_check(out(key), raw(path), checks, key) capture = json.loads(text_value(source.attrs.get("capture_metadata_json", "{}"))) # The mounting value is also present in FT config metadata. Do not infer # a sensor or TCP transform from it. config = json.loads(text_value(ft.attrs.get("config_json", "{}"))) mounting = capture.get("mounting") or config.get("mounting", {}) extension = mounting.get("axial_extension_m") if checks.require(isinstance(extension, (int, float)) and not isinstance(extension, bool) and np.isfinite(extension) and abs(extension - 0.063) < 1e-12, "recorded mounting extension must be 0.063 m"): checks.require(mounting.get("axis") == "eef_local_+z", "mounting offset axis is not eef_local_+z") for prefix in ("action", "observation/robot_state"): offset_check(raw(prefix + "/cartesian_position"), raw(prefix + "/cartesian_position_with_ft_offset"), float(extension), checks, prefix) unix_ns = raw("observation/timestamp/control/step_start_unix_ns") start_ns = int(source["observation/timestamp/control/step_start_unix_ns"][0]) checks.equal(out("timestamp_unix_ns"), unix_ns, "exact host timestamp") expected_times = (unix_ns - start_ns).astype(np.float64) / 1e9 checks.require(np.allclose(out("timestamps"), expected_times, atol=1e-12, rtol=0), "relative timestamps differ from native host step_start") checks.require(np.all(np.diff(out("timestamps")) > 0), "relative timestamps not strictly increasing") checks.require(np.all(np.diff(unix_ns) > 0), "host timestamps not strictly increasing") dispatched = raw("action/command_dispatched") skip = raw("observation/timestamp/skip_action") checks.equal(out("command_dispatched"), dispatched, "command dispatch extension") checks.equal(dispatched, ~skip, "command_dispatched must be inverse of skip_action") result["command_dispatched"] = int(np.count_nonzero(dispatched)) result["skipped_actions"] = int(np.count_nonzero(skip)) samples = ft["samples"] F = int(samples["wrench"].shape[0]) result["ft_samples_native"] = F checks.require(F > 0 and samples["wrench"].shape == (F, 6), "full FT wrench shape invalid") for key, obj in samples.items(): checks.require(obj.shape[0] == F, "FT sidecar length mismatch: " + key) checks.equal(samples["sample_index"][:], np.arange(F), "full FT sequence indices") checks.require(np.isfinite(samples["wrench"][:]).all(), "full FT stream contains nonfinite wrench") receive_ns = samples["receive_monotonic_ns"][:] request_ns = samples["request_monotonic_ns"][:] checks.require(np.all(np.diff(receive_ns) > 0), "full FT receive clock not strictly increasing") checks.require(np.all(request_ns <= receive_ns), "FT request occurs after receive") for attr in ("samples_received", "samples_written"): checks.require(int(ft.attrs.get(attr, -1)) == F, "FT " + attr + " differs from saved stream length") indices = raw("observation/ft_sensor/sample_index") if checks.require(indices.dtype.kind in "iu" and np.all(indices >= 0) and np.all(indices < F) and np.all(np.diff(indices) >= 0), "FT snapshot indices invalid"): for key in ("sample_index", "wrench", "packet_no", "packet_delta", "packet_discontinuity", "request_monotonic_ns", "request_unix_ns", "receive_monotonic_ns", "receive_unix_ns"): # Reading full small 100 Hz numeric arrays also allows repeated indices. checks.equal(raw("observation/ft_sensor/" + key), samples[key][:][indices], "FT snapshot " + key) checks.require(np.all(raw("observation/ft_sensor/valid")), "invalid FT snapshot exported") snapshot = raw("observation/ft_sensor/snapshot_monotonic_ns") received = raw("observation/ft_sensor/receive_monotonic_ns") checks.require(np.all(received <= snapshot), "FT snapshot uses future received sample") checks.require(np.all(snapshot <= raw("observation/timestamp/robot_state/read_start_monotonic_ns")), "FT snapshot follows robot read start") checks.require(np.allclose(raw("observation/ft_sensor/age_s"), (snapshot - received) / 1e9, atol=1e-12, rtol=0), "FT age does not match host monotonic clocks") wrench = raw("observation/ft_sensor/wrench") checks.equal(out("ft_wrench"), wrench, "full six-axis aligned FT") checks.equal(out("supplementary_tactile"), wrench, "supplementary six-axis FT") checks.equal(out("right_tactile_data_wrist_torque"), wrench[:, None, 3:], "standard wrist torque must be Mx/My/Mz only") camera_report = {} camera_audit = {c["serial"]: c for c in (audit or {}).get("replay_cameras", [])} for role, serial in CAMERAS.items(): variants = ((role + "_frame_index", "camera_frame_index__" + serial), (role + "_timestamp_ns", "camera_timestamp_unix_ns__" + serial)) values = [] for alternatives in variants: present = [key for key in alternatives if key in data] if not checks.require(bool(present), "missing camera provenance: " + "/".join(alternatives)): values.append(None) continue value = out(present[0]) checks.require(value.shape == (len(rows),) and value.dtype.kind in "iu", present[0] + ": expected integer vector") for other in present[1:]: checks.equal(value, out(other), "camera alias " + other) values.append(value) frame_indices, frame_times = values if frame_indices is not None: checks.require(np.all(frame_indices >= 0) and np.all(np.diff(frame_indices) > 0), role + ": invalid/repeated frame index") if frame_times is not None: checks.require(np.all(np.diff(frame_times) > 0), role + ": camera timestamps not strictly increasing") checks.equal(frame_times // 1_000_000, raw("observation/timestamp/cameras/" + serial + "_frame_received"), role + ": SDK timestamp must match original H5 milliseconds") known = camera_audit.get(serial) if known is not None and frame_indices is not None: N = int(known["decoded_frames"]) checks.require(np.all(frame_indices < N), role + ": frame index exceeds audited SVO count") # The frozen selection has two complete, H5-aligned T-1 prefixes. if N == raw_T - 1 and int(known.get("svo_timestamps_matched_to_trajectory", -1)) == N: checks.equal(rows, np.arange(N), role + ": not all audited real-image control rows retained") checks.equal(frame_indices, np.arange(N), role + ": real-image frame index mapping differs from audited prefix") if frame_times is not None: checks.require(int(frame_times[0]) == int(known["first_image_unix_ns"]), role + ": first exact SDK timestamp differs from audit") checks.require(int(frame_times[-1]) == int(known["last_image_unix_ns"]), role + ": last exact SDK timestamp differs from audit") rgb = data[role + "_rgb"] tested = 0 if scan_rgb_all: tested = scan_rgb_rows(rgb, span, checks, role) else: for pos in sorted({span.start, span.stop - 1}): frame = np.asarray(rgb[pos]) checks.require(frame.shape == (360, 640, 3) and frame.dtype == np.uint8, role + ": sampled RGB frame malformed") tested += 1 camera_report[serial] = {"role": role, "rows": len(rows), "rgb_rows_read": tested, "source_count_evidence": known is not None} result["cameras"] = camera_report if audit is not None: checks.require(int(audit["control_steps"]) == raw_T, "source audit native control count mismatch") checks.require(int(audit["ft_samples"]) == F, "source audit full FT count mismatch") result["checks"] = checks.count result["passed"] = not checks.errors return result def validate_export(root: Path, source_audit: Path | None = None, scan_rgb_all: bool = False) -> dict: import zarr root = root.resolve() checks = Checks() audit_path = find_audit(root, source_audit) audit_doc = json.loads(audit_path.read_text()) if audit_path else None audit_by_id = {e["episode_id"]: e for e in (audit_doc or {}).get("episodes", [])} result: dict[str, Any] = { "schema": "franka-ft-oxt-validation-v1", "dataset_root": str(root), "generated_at_utc": datetime.now(timezone.utc).isoformat(), "source_audit": str(audit_path) if audit_path else None, "errors": checks.errors, "tasks": {}, "scope": {"native_row_arrays": "all retained rows and all original row datasets compared exactly", "ft_stream": "all saved samples counted and checked; all aligned snapshots compared", "rgb": "all chunks read" if scan_rgb_all else "first and last retained frame of each episode read", "svo_decoding": "not repeated; existing source audit count/timestamp evidence used", "hash_verification": "separate process; not performed here", "physical_calibration": "not validated", "oxt_canonical_wrist_or_tcp_alignment": "not established; standard right_wrist_pose omitted, right_arm_joints provided; native flange observations retained as native__ fields", "gripper_slot_28": "normalized opening = 1 - native closure; 0=closed, 1=open; dimensionless"}, "limitations": ["This report establishes preservation and numeric consistency, not physical accuracy or final OXT acceptance.", "Native FT axes/origin and N/Nm unit confirmation remain as recorded; no TCP/base transform or software bias/gravity compensation is inferred.", "Logged camera extrinsics are historical calibration estimates; validity after FT mounting is unconfirmed."], } if audit_doc is None: result["limitations"].append("No source selection audit supplied; exact selected episode membership and SVO count evidence were not verified.") found = {p.stem for p in (root / "zarr").glob("*.zarr")} checks.require(found == set(TASKS), "expected exactly four task Zarr groups: " + ", ".join(TASKS)) exported_ids: list[str] = [] for task in TASKS: task_result: dict[str, Any] = {"episodes": [], "errors": []} result["tasks"][task] = task_result local = Checks() task_result["errors"] = local.errors try: group = zarr.open_group(str(root / "zarr" / (task + ".zarr")), mode="r") local.require(int(group.metadata.zarr_format) == 2, "expected Zarr format 2") data, meta = group["data"], group["meta"] ends = np.asarray(meta["episode_ends"][:]) if not local.require(ends.ndim == 1 and ends.dtype.kind in "iu" and len(ends) > 0 and ends[0] > 0 and np.all(np.diff(ends.astype(np.int64)) > 0), "episode_ends must be positive increasing integer vector"): continue T = int(ends[-1]) task_result["frames"] = T for key in data.array_keys(): local.require(data[key].ndim >= 1 and data[key].shape[0] == T, key + ": inconsistent first dimension") for key, tail in STANDARD_SHAPES.items(): local.require(key in data and data[key].shape == (T,) + tail, key + ": missing or wrong shape") local.require("right_wrist_pose" not in data, "right_wrist_pose must be omitted until a verified FTP-1 gripper-frame transform is available") for role in CAMERAS: key = role + "_rgb" local.require(key in data and data[key].shape == (T, 360, 640, 3) and data[key].dtype == np.uint8, key + ": expected uint8(T,360,640,3)") for key in ("episode_ids", "episode_names", "source_episode_paths"): local.require(key in meta and meta[key].shape == ends.shape, "meta/" + key + ": missing or wrong shape") if local.errors: continue local.require(np.all(data["right_hand_joints_idx"][:] == 28), "gripper slot must be 28") local.require(data["right_hand_joints_idx"].dtype.kind in "iu", "gripper slot dtype must be integer") local.require(np.all(data["right_tactile_area_wrist_torque"][:] == 15), "wrist torque area must be 15") local.require(data["right_tactile_area_wrist_torque"].dtype.kind in "iu", "wrist torque area dtype must be integer") local.require(data["timestamp_unix_ns"].dtype == np.dtype("int64"), "timestamp_unix_ns must preserve int64 precision") local.require(data["command_dispatched"].dtype == np.dtype("bool"), "command_dispatched must retain boolean dtype") local.require(np.all(strings(data["right_tactile_type_wrist_torque"][:]) == "state"), "wrist torque type must be state") local.require(np.all(strings(data["right_tactile_sensor_wrist_torque"][:]) == "SRIM3813A1"), "unexpected wrist torque sensor identifier") instructions = strings(data["sub_task_instruction"][:]) local.require(bool(np.all(np.char.str_len(np.char.strip(instructions)) > 0)), "empty task instruction") for key in STANDARD_SHAPES: if data[key].dtype.kind in "fc": local.require(np.isfinite(data[key][:]).all(), key + ": nonfinite standard values") ids = strings(meta["episode_ids"][:]).tolist() names = strings(meta["episode_names"][:]).tolist() paths = strings(meta["source_episode_paths"][:]).tolist() local.require(len(ids) == len(set(ids)), "duplicate episode IDs within task") local.require(len(paths) == len(set(paths)), "duplicate native episode paths within task") start = 0 for index, end in enumerate(ends): episode_id = ids[index] exported_ids.append(episode_id) known = audit_by_id.get(episode_id) if audit_doc is not None: local.require(known is not None, "episode absent from source audit: " + episode_id) if known is not None: expected_task = "insert_pen" if known["task_slug"] == "insert" else known["task_slug"] local.require(expected_task == task, "episode assigned to wrong task: " + episode_id) try: ep = validate_episode(data, slice(start, int(end)), root, paths[index], episode_id, names[index], known, scan_rgb_all) except Exception as exc: ep = {"episode_id": episode_id, "episode_name": names[index], "passed": False, "errors": [f"{type(exc).__name__}: {exc}"]} task_result["episodes"].append(ep) start = int(end) task_result["checks"] = local.count except Exception as exc: local.errors.append(f"{type(exc).__name__}: {exc}") finally: task_result["passed"] = not local.errors and bool(task_result["episodes"]) and all(e.get("passed", False) for e in task_result["episodes"]) checks.require(len(exported_ids) == len(set(exported_ids)), "duplicate episode IDs across tasks") if audit_doc is not None: checks.require(set(exported_ids) == set(audit_by_id), "exported episode set differs from frozen source audit") episodes = [e for task in result["tasks"].values() for e in task["episodes"]] result["summary"] = { "tasks": len(result["tasks"]), "episodes": len(episodes), "standard_frames": sum(t.get("frames", 0) for t in result["tasks"].values()), "native_control_rows": sum(e.get("control_rows_native", 0) for e in episodes), "omitted_standard_control_rows": sum(e.get("control_rows_omitted", 0) for e in episodes), "native_ft_samples": sum(e.get("ft_samples_native", 0) for e in episodes), "failed_episodes": sum(not e.get("passed", False) for e in episodes), } result["passed"] = not checks.errors and all(t["passed"] for t in result["tasks"].values()) return result def main(argv: list[str] | None = None) -> int: parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--dataset-root", type=Path, required=True) parser.add_argument("--report", type=Path, required=True) parser.add_argument("--source-audit", type=Path) parser.add_argument("--scan-rgb-all", action="store_true", help="Read all exported RGB chunks instead of per-episode endpoints") args = parser.parse_args(argv) try: report = validate_export(args.dataset_root, args.source_audit, args.scan_rgb_all) except Exception as exc: report = {"schema": "franka-ft-oxt-validation-v1", "passed": False, "errors": [f"{type(exc).__name__}: {exc}"]} args.report.parent.mkdir(parents=True, exist_ok=True) temporary = args.report.with_name(args.report.name + ".partial") temporary.write_text(json.dumps(report, ensure_ascii=False, indent=2) + "\n") temporary.replace(args.report) print(json.dumps({"passed": report["passed"], "report": str(args.report), "summary": report.get("summary"), "errors": report["errors"]}, ensure_ascii=False)) return 0 if report["passed"] else 1 if __name__ == "__main__": sys.exit(main())