Blame for code/case1/runner/art_runner/config.py

"""Episode configuration: validation, defaults, canonical hash (NEW CODE).

The authoritative schema is runner/schemas/episode_config.schema.json; this
module implements the same rules without third-party dependencies. Unknown
keys are rejected so that typos cannot silently fall back to defaults.
"""
import copy
import hashlib
import json
import math
import re

DT = 1e-3
SCHEMA_VERSION = 'tr-c1-episode/1'
RELEASED = ('mpc', 'nnmpc', 'pid', 'nnhd')

DEFAULT_PERTURBATION = {
    'steering_bias': 0.0,
    'actuation_delay_s': 0.0,
    'friction': None,
    'obs_noise': {'seed': 0, 'std_x': 0.0, 'std_y': 0.0, 'std_yaw': 0.0},
}
DEFAULT_TERMINATION = {
    'type': 'fixed_horizon',
    'tube_lat_m': 0.1,
    'tube_head_rad': 0.1,
    'divergence_dist_m': 5.0,
    'divergence_hold_s': 2.0,
    'stop_on_divergence': True,
    'completion_margin_s': 0.0,
}
ROS_DEFAULTS = {'controller_period_s': 0.1, 'state_rate_hz': 25.0, 'driver_rate_hz': 10.0, 'state_delay_s': 0.0}
TOP_KEYS = {'schema', 'episode_id', 'scenario_id', 'controller', 'controller_params', 'reference', 'initial_pose',
            'horizon_s', 'control_start_s', 'latency', 'perturbation', 'termination', 'on_controller_failure',
            'replicate_waypoint_boxes', 'replicate_visual_setters', 'tags', 'mpc_osqp_adaptive_rho_interval'}
DEFAULT_OSQP_ADAPTIVE_RHO_INTERVAL = 25


class ConfigError(ValueError):
    pass


# Parameters and DEFAULTS of
# every plug-in. config.normalize rejects an unknown plug-in name, an unknown parameter (typos cannot silently fall
# back to a default) and a non-finite / out-of-range value. normalize() does NOT write the defaults into the config
# (configs written before this revision keep their canonical hash); the plug-in applies them and the episode summary
# records them as `controller_params_effective`. Kinds: num = finite number, pos = finite > 0, nonneg = finite >= 0,
# bool, released = one of RELEASED. REQUIRED = no default.
REQUIRED = object()
PLUGIN_PARAM_SPECS = {
    'replica_pid': {},
    # pure pursuit / Stanley: P speed loop throttle = clip(throttle_ff + k_v * (v_ref - v), 0, 1)
    'pure_pursuit': {'lookahead_m': ('pos', 1.0), 'wheelbase_m': ('pos', 0.5), 'k_v': ('num', 1.0),
                     'throttle_ff': ('num', 0.0)},
    'stanley': {'k': ('num', 1.0), 'softening': ('nonneg', 0.5), 'k_v': ('num', 1.0), 'throttle_ff': ('num', 0.0)},
    'wrapped_released': {'base': ('released', REQUIRED), 'steering_scale': ('num', 1.0), 'shared_speed': ('bool', False),
                         'kp_speed': ('num', 1.0), 'throttle_bias': ('num', 0.0)},
    'mpc_n5': {},
}


def plugin_effective_params(name, params):
    """Validate plug-in parameters against PLUGIN_PARAM_SPECS and return them with the defaults filled in."""
    if name not in PLUGIN_PARAM_SPECS:
        raise ConfigError(f'unknown plug-in {name!r}; known plug-ins: {sorted(PLUGIN_PARAM_SPECS)}')
    spec = PLUGIN_PARAM_SPECS[name]
    if spec is None:  # test-only registrations (fault-injection checks) skip validation
        return dict(params or {})
    params = dict(params or {})
    _check_keys(params, spec, f'controller_params of plugin:{name}')
    out = {}
    for k, (kind, default) in spec.items():
        if k not in params:
            if default is REQUIRED:
                raise ConfigError(f'plugin:{name} requires controller_params.{k}')
            out[k] = default
            continue
        v = params[k]
        if kind == 'bool':
            if not isinstance(v, bool):
                raise ConfigError(f'plugin:{name} {k} must be true/false')
        elif kind == 'released':
            if v not in RELEASED:
                raise ConfigError(f'plugin:{name} {k} must be one of {RELEASED}')
        else:
            _num(params, k, lo=None)
            if kind == 'pos' and not v > 0:
                raise ConfigError(f'plugin:{name} {k} must be > 0')
            if kind == 'nonneg' and v < 0:
                raise ConfigError(f'plugin:{name} {k} must be >= 0')
        out[k] = v
    return out


def _num(d, k, lo=None, hi=None, required=True, default=None):
    if k not in d:
        if required:
            raise ConfigError(f'missing {k}')
        return default
    v = d[k]
    if isinstance(v, bool) or not isinstance(v, (int, float)) or not math.isfinite(v):
        raise ConfigError(f'{k} must be a finite number, got {v!r}')
    if lo is not None and v < lo:
        raise ConfigError(f'{k}={v} < {lo}')
    if hi is not None and v > hi:
        raise ConfigError(f'{k}={v} > {hi}')
    return float(v)


def _check_keys(d, allowed, where):
    extra = set(d) - set(allowed)
    if extra:
        raise ConfigError(f'unknown keys in {where}: {sorted(extra)}')


def to_steps(seconds, name):
    n = seconds / DT
    r = int(round(n))
    if abs(n - r) > 1e-6:
        raise ConfigError(f'{name}={seconds} s is not a multiple of the 1 ms physics step')
    return r


def normalize(cfg):
    if not isinstance(cfg, dict):
        raise ConfigError('episode config must be an object')
    c = copy.deepcopy(cfg)
    _check_keys(c, TOP_KEYS, 'episode')
    c['schema'] = c.get('schema', SCHEMA_VERSION)
    if c['schema'] != SCHEMA_VERSION:
        raise ConfigError(f"schema {c['schema']} != {SCHEMA_VERSION}")
    for k in ('episode_id', 'scenario_id'):
        if not isinstance(c.get(k), str) or not re.fullmatch(r'[A-Za-z0-9_.\-]{1,160}', c[k]):
            raise ConfigError(f'{k} must match [A-Za-z0-9_.-]{{1,160}}')
    ctrl = c.get('controller')
    if not isinstance(ctrl, str) or not (ctrl in RELEASED or ctrl.startswith('plugin:')):
        raise ConfigError(f'controller must be one of {RELEASED} or plugin:<name>')
    c['controller_params'] = c.get('controller_params', {})
    if ctrl in RELEASED and c['controller_params']:
        raise ConfigError('released controllers take no controller_params (unmodified)')
    if ctrl.startswith('plugin:'):
        if not isinstance(c['controller_params'], dict):
            raise ConfigError('controller_params must be an object')
        plugin_effective_params(ctrl.split(':', 1)[1], c['controller_params'])  # validation only
    ref = c.get('reference')
    if not isinstance(ref, dict) or ref.get('type') not in ('released_csv', 'csv_transformed', 'straight', 'array'):
        raise ConfigError('reference.type must be released_csv|csv_transformed|straight|array')
    allowed_ref = {'released_csv': {'type', 'name'}, 'csv_transformed': {'type', 'name', 'dx', 'dy', 'rotate'},
                   'straight': {'type', 'x0', 'y0', 'heading', 's_start', 's_end', 'spacing', 'v_ref'},
                   'array': {'type', 'points'}}[ref['type']]
    _check_keys(ref, allowed_ref, 'reference')
    if ref['type'] in ('released_csv', 'csv_transformed') and not re.fullmatch(r'[A-Za-z0-9_]+', str(ref.get('name'))):
        raise ConfigError('reference.name must be a released waypoint stem')
    pose = c.get('initial_pose')
    if not isinstance(pose, dict):
        raise ConfigError('initial_pose required')
    _check_keys(pose, {'x', 'y', 'yaw', 'frame'}, 'initial_pose')
    for k in ('x', 'y', 'yaw'):
        _num(pose, k)
    pose['frame'] = pose.get('frame', 'chassis_ref')
    if pose['frame'] not in ('chassis_ref', 'cog'):
        raise ConfigError('initial_pose.frame must be chassis_ref|cog')
    _num(c, 'horizon_s', lo=0.001, hi=3600)
    to_steps(c['horizon_s'], 'horizon_s')
    c['control_start_s'] = _num(c, 'control_start_s', lo=0, required=False, default=0.0)
    to_steps(c['control_start_s'], 'control_start_s')
    lat = c.get('latency')
    if not isinstance(lat, dict) or lat.get('model') not in ('ideal', 'ros_emulated'):
        raise ConfigError('latency.model must be given explicitly: ideal|ros_emulated (no campaign default)')
    if lat['model'] == 'ideal':
        _check_keys(lat, {'model', 'controller_period_s'}, 'latency')
        lat['controller_period_s'] = _num(lat, 'controller_period_s', lo=0.001, required=False, default=0.1)
        to_steps(lat['controller_period_s'], 'controller_period_s')
    else:
        _check_keys(lat, {'model', 'controller_phase_s', 'compute_delay_s', 'state_delay_s', 'controller_period_s',
                          'state_rate_hz', 'driver_rate_hz'}, 'latency')
        for k in ('controller_phase_s', 'compute_delay_s'):
            _num(lat, k, lo=0)  # required: calibrated by the ROS-reference agent
            to_steps(lat[k], k)
        for k, v in ROS_DEFAULTS.items():
            lat[k] = _num(lat, k, lo=0, required=False, default=v)
        to_steps(lat['state_delay_s'], 'state_delay_s')
        to_steps(lat['controller_period_s'], 'controller_period_s')
    pert = copy.deepcopy(DEFAULT_PERTURBATION)
    user_p = c.get('perturbation', {})
    _check_keys(user_p, set(DEFAULT_PERTURBATION), 'perturbation')
    for k in ('steering_bias', 'actuation_delay_s'):
        if k in user_p:
            pert[k] = _num(user_p, k, lo=-1 if k == 'steering_bias' else 0, hi=1 if k == 'steering_bias' else 10)
    to_steps(pert['actuation_delay_s'], 'actuation_delay_s')
    if user_p.get('friction') is not None:
        pert['friction'] = _num(user_p, 'friction', lo=0.0, hi=2.0)
    if 'obs_noise' in user_p:
        on = user_p['obs_noise']
        _check_keys(on, {'seed', 'std_x', 'std_y', 'std_yaw'}, 'perturbation.obs_noise')
        for k in ('std_x', 'std_y', 'std_yaw'):
            if k in on:
                pert['obs_noise'][k] = _num(on, k, lo=0)
        if 'seed' in on:
            if not isinstance(on['seed'], int) or isinstance(on['seed'], bool) or on['seed'] < 0:
                raise ConfigError('obs_noise.seed must be a non-negative integer')
            pert['obs_noise']['seed'] = on['seed']
    c['perturbation'] = pert
    term = copy.deepcopy(DEFAULT_TERMINATION)
    user_t = c.get('termination', {})
    _check_keys(user_t, set(DEFAULT_TERMINATION), 'termination')
    if 'type' in user_t:
        if user_t['type'] not in ('fixed_horizon', 'path_completion'):
            raise ConfigError('termination.type must be fixed_horizon|path_completion')
        term['type'] = user_t['type']
    for k in ('tube_lat_m', 'tube_head_rad', 'divergence_dist_m', 'divergence_hold_s', 'completion_margin_s'):
        if k in user_t:
            term[k] = _num(user_t, k, lo=0)
    if 'stop_on_divergence' in user_t:
        term['stop_on_divergence'] = bool(user_t['stop_on_divergence'])
    to_steps(term['divergence_hold_s'], 'divergence_hold_s')
    to_steps(term['completion_margin_s'], 'completion_margin_s')
    c['termination'] = term
    c['on_controller_failure'] = c.get('on_controller_failure', 'hold_last_command')
    if c['on_controller_failure'] not in ('hold_last_command', 'terminate'):
        raise ConfigError('on_controller_failure must be hold_last_command|terminate')
    # The released MPC leaves OSQP's adaptive_rho_interval at 0 ("automatic"), which OSQP 0.6 resolves from
    # WALL-CLOCK time (first iteration count at which solve time > 0.4 x setup time, rounded to a multiple of
    # check_termination=25). Under CPU load this changes solutions within the solver tolerance. 25 is the value
    # that reproduces every archived nominal-sim MPC command exactly (checked by replay);
    # null keeps the released time-based behaviour (NOT deterministic).
    v = c.get('mpc_osqp_adaptive_rho_interval', DEFAULT_OSQP_ADAPTIVE_RHO_INTERVAL)
    if v is not None and (not isinstance(v, int) or isinstance(v, bool) or v <= 0):
        raise ConfigError('mpc_osqp_adaptive_rho_interval must be a positive integer or null')
    c['mpc_osqp_adaptive_rho_interval'] = v
    c['replicate_waypoint_boxes'] = bool(c.get('replicate_waypoint_boxes', False))
    c['replicate_visual_setters'] = bool(c.get('replicate_visual_setters', False))
    c['tags'] = c.get('tags', {})
    return c


def config_hash(normalized):
    blob = json.dumps(normalized, sort_keys=True, separators=(',', ':'), allow_nan=False)
    return hashlib.sha256(blob.encode()).hexdigest()


def load_episode_list(path):
    with open(path) as f:
        data = json.load(f)
    if isinstance(data, dict):
        data = data.get('episodes')
    if not isinstance(data, list) or not data:
        raise ConfigError('episode file must be a non-empty list or {"episodes": [...]}')
    out, seen = [], set()
    for i, cfg in enumerate(data):
        try:
            n = normalize(cfg)
        except ConfigError as e:
            raise ConfigError(f'episode #{i} ({cfg.get("episode_id") if isinstance(cfg, dict) else "?"}): {e}')
        if n['episode_id'] in seen:
            raise ConfigError(f"duplicate episode_id {n['episode_id']}")
        seen.add(n['episode_id'])
        out.append(n)
    return out