What causes "too few samples found in route" in commaai/openpilot?

Record a longer route with openpilot engaged and hands off (no driver torque) so >512 consecutive valid samples accumulate.

commaai/openpilot · error

too few samples found in route

Error message

too few samples found in route

What it means

Raised at the end of sample collection in toyota_eps_factor.py when fewer than MIN_SAMPLES (512) paired (torque_cmd, eps_torque) samples were gathered from the route's CAN data. The script needs a continuous engaged run with LKA torque commands and matching EPS motor torque readings while the driver is not overriding; if the route never accumulates that many valid pairs, the linear regression cannot be fitted reliably.

Source

Thrown at tools/scripts/car/toyota_eps_factor.py:46

    for m in msg.can:
      if m.address == 0x2e4 and m.src == 128:
        engaged = bool(m.dat[0] & 1)
        torque_cmd = to_signed((m.dat[1] << 8) | m.dat[2], 16)
      elif m.address == 0x260 and m.src == 0:
        eps_torque = to_signed((m.dat[5] << 8) | m.dat[6], 16)
        steering_pressed = abs(to_signed((m.dat[1] << 8) | m.dat[2], 16)) > STEER_THRESHOLD

    if engaged and torque_cmd is not None and eps_torque is not None and not steering_pressed:
      cmds.append(torque_cmd)
      eps.append(eps_torque)
    else:
      if len(cmds) > MIN_SAMPLES:
        break
      cmds, eps = [], []

  if len(cmds) < MIN_SAMPLES:
    raise Exception("too few samples found in route")

  lm = linear_model.LinearRegression(fit_intercept=False)
  lm.fit(np.array(cmds).reshape(-1, 1), eps)
  scale_factor = 1. / lm.coef_[0]

  if plot:
    plt.plot(np.array(eps) * scale_factor)
    plt.plot(cmds)
    plt.show()
  return scale_factor


if __name__ == "__main__":
  lr = LogReader(sys.argv[1])
  n = get_eps_factor(lr, plot="--plot" in sys.argv)
  print("EPS torque factor: ", n)

View on GitHub (pinned to 516ec1e682)

Solutions

Record a longer route with openpilot engaged and hands off (no driver torque) so >512 consecutive valid samples accumulate.
Verify the car actually sends the torque messages this script decodes (check the CAN IDs at the top of toyota_eps_factor.py against your car's bus logs).
Lower MIN_SAMPLES in the script only as a last resort — the fitted EPS factor will be noisier and affect steer accuracy.
Make sure you're passing a route with steering engaged segments, not just a drive with openpilot passive.

Example fix

# before
# short 30s drive, driver hands on -> too few samples
python tools/scripts/car/toyota_eps_factor.py <route>

# after
# record several minutes engaged, hands off, gentle curves, then:
python tools/scripts/car/toyota_eps_factor.py <long_engaged_route>

Defensive patterns

Strategy: validation

Try / catch

try:
    factor = toyota_eps_factor(route, MIN_SAMPLES=512)
except Exception as e:
    if 'too few samples' in str(e):
        # collect a longer engaged, hands-off route before retrying
        sys.exit('need a longer engaged drive (>512 samples)')
    raise

Prevention

Record multi-minute routes with openpilot engaged and no driver torque on the wheel.
Confirm the car emits the Toyota torque CAN IDs the script decodes before collecting data.
Don't lower MIN_SAMPLES without increasing the measurement noise budget for the EPS factor.

When it happens

Trigger: Processing a route where the car was never openpilot-engaged long enough, where STEER_THRESHOLD-qualifying driver torque kept resetting the sample buffer (any steering_pressed event clears cmds/eps), or where the expected Toyota CAN IDs for torque_cmd/eps_torque never appeared (wrong car or DBC).

Common situations: Short test drives, routes where the driver kept touching the wheel, wrong car model (no TOYOTA torque messages on those IDs), or a segment with gaps in CAN data. Users frequently hit this on a first quick recording before doing a proper engaged drive.

Related errors

AI-assisted analysis of commaai/openpilot@516ec1e682 (2026-08-15). Data as JSON: /api/errors/9b4d113d57359ddd. Report an issue: GitHub.