Files
ros_contorl/scripts/run_handeye.sh
T
arm_dev 2b77ec127d P4:权重/标签模块 —— 旁挂标签文件 + PT→ONNX→TRT 转换工具 + 归档归位
痛点:TensorRT 10.13 的 Python API 没有元数据读写,engine 一转就把类名丢了;
推理侧只能退化成 str(cid)("0"、"1"),界面上看不出是什么类,属于静默失效。
同时旧的 export_yolo_trt.sh 只能手工搬文件、手工改 site.yaml、标签无处安放。

标签子系统(model_labels.py)
- 旁挂标签文件 <模型名>.labels.yaml,放在权重旁边,文本、可手改、进 git
  (二进制权重不进 git,标签是配置所以进)
- 读取优先级:旁挂文件 > ONNX 元数据 > 模型元数据 > 占位符 class_N
  占位符是故意的:缺标签时要一眼看出来,而不是悄悄退回 cid
- 解析器容忍多种写法:dict / list / 逗号串 / 逐行文本 / Python repr 串

布局与转换(weights_layout.py + weights.py + weights.sh)
- models/<分组>/<模型名>.{pt,onnx,engine} + 同名 .labels.yaml
- convert:PT→ONNX→TensorRT,支持自定义名字、自动归位、固化标签
- staging/:非标准格式先进这里,转换后源文件归档到 staging/done/
- list / labels / doctor / show 四个子命令
- doctor --check-engines 能查出 engine 与 TensorRT 版本不匹配(反序列化返回空)

推理侧接入
- yolo_detector / onnx_workpiece / hole_detect_hough 统一走 read_labels()
- 画框标签不再硬编码 "gongjian",改用实际类名

顺手修掉的真实问题
- setup.py 撞上断掉的符号链接会让整包构建失败,报错还指向 build/ 里的幽灵文件
  (源文件被挪走后 build/ 里留下的断链被 os.walk 列出来)。加 _existing() 过滤。
- P3 遗漏:site.yaml / export_site_env.py / run_handeye.sh / handeye_teach.py /
  handeye_calibrate.py 里的 handeye 工作目录仍是 /tmp/handeye_calib*,
  与已迁到 handeye/ 的其余路径不一致。统一到 handeye/tmp/<臂号>。
- site.yaml 的 hole_weights / hole_detect.weights 原指向不存在的 best.pt,
  改为实际存在的 detect_best_1.onnx;active.pt 符号链接已废弃并清理。
- weights.sh labels 原先不接受 "分组/模型名" 写法,而 list 输出正是这个形式,
  照着敲会报"找不到模型"。补上,并让报错直接给出可复制的全名。

测试:225 项通过(新增 test_model_labels.py / test_weights_layout.py)。
真实 ONNX→engine 端到端转换已验过。

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-23 12:39:18 +08:00

381 lines
14 KiB
Bash
Executable File
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
#!/usr/bin/env bash
# Hand-eye calibration helpers.
# Scripts: ~/ros2/scripts/run_handeye.sh
# Config & results: ~/ros2/handeye/ ← 见 handeye/README.md
# poses/<臂号>_<模式>.yaml 采点(输入)
# results/<臂号>_<模式>.yaml 标定结果(输出)
# calibrate.yaml 板尺寸 / 运动 / 稳定性参数
# tmp/ 临时产物
# Site overrides: ~/ros2/config/site.yaml → handeye:
#
# Usage:
# bash ~/ros2/scripts/run_handeye.sh calibrate
# bash ~/ros2/scripts/run_handeye.sh waypoints
# bash ~/ros2/scripts/run_handeye.sh teach # RDK 软拖拽(推荐)
# bash ~/ros2/scripts/run_handeye.sh teach --ros --auto-solve
# # 示教器模式 + 采完自动 merge 并直接标定(一条命令跑完)
# bash ~/ros2/scripts/run_handeye.sh teach --points 12 # 目标采 12 个点
# bash ~/ros2/scripts/run_handeye.sh teach --append # 接着上次点位继续采
# bash ~/ros2/scripts/run_handeye.sh teach --ros # 示教器 Free Drive
# bash ~/ros2/scripts/run_handeye.sh merge [taught.txt]
# bash ~/ros2/scripts/run_handeye.sh show # 看当前配置与点数
#
# 关于 --auto-solve(采完自动标定)的重要限制:
# 只有 --ros(示教器)模式能真正一条命令跑完。因为标定必须在 run_real.sh 的
# 控制器上走点位,而 --rdk 软拖拽要 RDK 独占,两者互斥 —— 这是硬件约束。
# --rdk --auto-solve 会照常自动 merge,然后明确告诉你接下来那两条命令。
#
# 全局选项(放在子命令前或后都可以):
# --arm 89|91|left|right 选臂(默认 89)
# --mode eye_in_hand|eye_to_hand
# 选模式。会切换到 <臂号>_<模式>.yaml,
# 所以 eye_in_hand 与 eye_to_hand 的结果互不覆盖
# --points N 仅 teach:目标采集点数(0=不限)
#
# 例:
# bash ~/ros2/scripts/run_handeye.sh --arm 91 teach --mode eye_to_hand --points 15
#
# calibrate / waypoints 需要:
# T1: bash ~/ros2/scripts/run_real.sh --arm 89|91
# T2: bash ~/ros2/scripts/run_camera_d405.sh --arm 89|91 --preview
# teach(拖拽)需要: 先停掉 run_real.sh 和相机(RDK + D405 都独占)
# 89 与 91 的结果文件、示教点、workdir 完全分开,禁止混用。
set -eo pipefail
ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
# shellcheck disable=SC1091
source "${ROOT}/scripts/lib/parse_arm_args.sh"
parse_arm_args "$@"
set -- "${PARSE_ARM_REST[@]}"
CMD="${1:-}"
shift || true
usage() {
sed -n '2,40p' "$0"
}
# shellcheck disable=SC1091
source "${ROOT}/scripts/env.sh"
# shellcheck disable=SC1091
source "${ROOT}/scripts/lib/load_site_config.sh"
# Prefer package source tree (editable); fall back to install share
APP_CFG="${ROOT}/src/arm_application/config"
if [[ ! -d "${APP_CFG}" ]]; then
APP_CFG="$(ros2 pkg prefix arm_application 2>/dev/null)/share/arm_application/config" || true
fi
# 手眼标定相关文件现在集中在仓库根的 handeye/ 下(见 handeye/README.md)。
# 这些 SHE 值正常情况下由 site.yaml → export_site_env.py 提供;
# 下面只是 shell 侧的兜底,路径必须与 export_site_env.py 的默认值一致。
HE_DIR="${SITE_HANDEYE_DIR:-${ROOT}/handeye}"
WP="${SITE_HANDEYE_WAYPOINTS:-${HE_DIR}/poses/89_eye_in_hand.yaml}"
CAL="${SITE_HANDEYE_CALIBRATE:-${HE_DIR}/calibrate.yaml}"
TOUR="${SITE_HANDEYE_TOUR:-${HE_DIR}/waypoints.yaml}"
RES="${SITE_HANDEYE_FILE:-${HE_DIR}/results/89_eye_in_hand.yaml}"
WORKDIR="${SITE_HANDEYE_WORKDIR:-${HE_DIR}/tmp}"
TAUGHT="${WORKDIR}/taught_poses.txt"
PREFIX="${SITE_JOINT_PREFIX:-Rizon4-063989_}"
CAM="${SITE_CAM_NAME:-d405}"
# ---- --mode:切换 eye-in-hand / eye-to-hand ----
# 必须在 WP/RES 定下来之前处理,因为它会改变读写的文件名。
# 其余参数原样留给子命令(teach 自己有 --rdk/--ros 等)。
HANDEYE_MODE="${SITE_HANDEYE_MODE:-eye_in_hand}"
_rest=()
while [[ $# -gt 0 ]]; do
case "$1" in
--mode)
HANDEYE_MODE="${2:?--mode 需要 eye_in_hand 或 eye_to_hand}"
if [[ "${HANDEYE_MODE}" != "eye_in_hand" && "${HANDEYE_MODE}" != "eye_to_hand" ]]; then
echo "ERROR: --mode 只能是 eye_in_hand 或 eye_to_hand,收到 '${HANDEYE_MODE}'" >&2
exit 1
fi
shift 2
;;
*)
_rest+=("$1")
shift
;;
esac
done
set -- "${_rest[@]}"
# 臂号取 Flexiv 序列号后两位:Rizon4-063989 → 89
_arm_id="${SITE_ROBOT_SN: -2}"
if [[ -n "${_arm_id}" && "${_arm_id}" != "${SITE_ROBOT_SN}" ]]; then
if [[ "${HANDEYE_MODE}" != "${SITE_HANDEYE_MODE:-eye_in_hand}" ]]; then
# 显式指定了非默认模式 → 按 <臂号>_<模式>.yaml 重新定位
WP="${HE_DIR}/poses/${_arm_id}_${HANDEYE_MODE}.yaml"
RES="${HE_DIR}/results/${_arm_id}_${HANDEYE_MODE}.yaml"
echo "[handeye] --mode ${HANDEYE_MODE} → poses: ${WP}"
echo "[handeye] result: ${RES}"
fi
fi
_kill_servo() {
pkill -f 'visual_servo.launch' 2>/dev/null || true
pkill -f 'lib/arm_application/servo_node' 2>/dev/null || true
pkill -f 'lib/arm_application/perception_node' 2>/dev/null || true
sleep 1
}
_need_files() {
local f
for f in "$@"; do
if [[ ! -f "$f" ]]; then
echo "ERROR: missing $f" >&2
exit 2
fi
done
}
_need_robot() {
if ! ros2 service list 2>/dev/null | grep -q '/set_named_pose'; then
echo "ERROR: start run_real.sh first" >&2
exit 1
fi
}
_need_camera() {
if ! ros2 topic list 2>/dev/null | grep -q "/${CAM}/color/image_raw"; then
echo "ERROR: no /${CAM}/color/image_raw — start run_camera_d405.sh" >&2
exit 1
fi
}
cmd_show() {
echo "=== hand-eye (package + site) ==="
echo " arm : ${SITE_ROBOT_SN:-} ${SITE_ROBOT_LABEL:-}"
echo " package cfg : ${APP_CFG}"
echo " calibrate : ${CAL}"
echo " poses : ${WP}"
echo " tour : ${TOUR}"
echo " result : ${RES}"
echo " workdir : ${WORKDIR}"
echo " board : ${SITE_HANDEYE_BOARD_COLS:-?}x${SITE_HANDEYE_BOARD_ROWS:-?} inner, square=${SITE_HANDEYE_SQUARE_M:-?} m"
python3 "${ROOT}/scripts/lib/export_site_env.py" --print || true
if [[ -f "${WP}" ]]; then
python3 - <<PY
import re
from pathlib import Path
t = Path("${WP}").read_text(encoding="utf-8")
m = re.search(r"num_poses:\s*(\d+)", t)
n = int(m.group(1)) if m else "?"
print(f" num_poses in poses yaml: {n}")
PY
fi
}
cmd_merge() {
local src="${1:-${TAUGHT}}"
_need_files "${src}" "${WP}"
python3 "${ROOT}/scripts/merge_handeye_poses.py" "${src}" "${WP}"
echo "=== next: bash ~/ros2/scripts/run_handeye.sh calibrate ==="
}
cmd_calibrate() {
_need_files "${CAL}" "${WP}"
_need_robot
_need_camera
echo "=== kill visual_servo (avoid traj fight) ==="
_kill_servo
mkdir -p "${WORKDIR}"
local cols="${SITE_HANDEYE_BOARD_COLS:-8}"
local rows="${SITE_HANDEYE_BOARD_ROWS:-11}"
local sq="${SITE_HANDEYE_SQUARE_M:-0.015}"
local vel="${SITE_HANDEYE_VEL_SCALE:-0.05}"
local mt="${SITE_HANDEYE_MOVE_T:-8.0}"
local st="${SITE_HANDEYE_SETTLE_SEC:-1.5}"
local qd="${SITE_HANDEYE_SETTLE_QD:-0.015}"
local ss="${SITE_HANDEYE_SETTLE_STABLE:-0.6}"
local sto="${SITE_HANDEYE_SETTLE_TO:-8.0}"
local ph="${SITE_HANDEYE_PIN_HOLDS:-2}"
local ps="${SITE_HANDEYE_PIN_SEC:-0.45}"
local skip="${SITE_HANDEYE_SKIP_UNSTABLE:-true}"
echo "=== calibrate mode=${HANDEYE_MODE} ==="
echo " board ${cols}x${rows} inner, square=${sq} m (from site.yaml)"
echo " motion v=${vel} T≥${mt}s settle=${st}s qd≤${qd} pin=${ph}x${ps}s"
echo " poses=${WP}"
echo " result → ${RES}"
exec ros2 run arm_application handeye_calibrate --ros-args \
--params-file "${CAL}" \
--params-file "${WP}" \
-p joint_prefix:="${PREFIX}" \
-p camera_source:="${CAM}" \
-p output_dir:="${WORKDIR}" \
-p result_file:="${RES}" \
-p mode:="${HANDEYE_MODE}" \
-p board_cols:="${cols}" \
-p board_rows:="${rows}" \
-p square_size_m:="${sq}" \
-p velocity_scale:="${vel}" \
-p move_time_sec:="${mt}" \
-p settle_sec:="${st}" \
-p settle_qd_max:="${qd}" \
-p settle_stable_sec:="${ss}" \
-p settle_timeout_sec:="${sto}" \
-p pin_holds:="${ph}" \
-p pin_hold_sec:="${ps}" \
-p skip_unstable_poses:="${skip}"
}
cmd_waypoints() {
_need_files "${TOUR}" "${WP}"
_need_robot
_need_camera
echo "=== kill visual_servo ==="
_kill_servo
mkdir -p "${WORKDIR}"
echo "=== waypoint tour (Enter=confirm, q=quit) ==="
echo " saves confirmed joints → ${TAUGHT}"
echo " then: bash ~/ros2/scripts/run_handeye.sh merge"
exec ros2 run arm_application handeye_waypoints --ros-args \
--params-file "${TOUR}" \
--params-file "${WP}" \
-p joint_prefix:="${PREFIX}" \
-p camera_source:="${CAM}" \
-p append_poses_file:="${TAUGHT}"
}
cmd_teach() {
local backend="rdk"
local append=false
local points=0
local auto_solve=false
while [[ $# -gt 0 ]]; do
case "$1" in
--rdk) backend="rdk"; shift ;;
--ros|--pendant) backend="ros"; shift ;;
--append|--keep) append=true; shift ;;
--points) points="${2:?--points 需要一个数字}"; shift 2 ;;
--auto-solve) auto_solve=true; shift ;;
*) echo "Unknown teach arg: $1" >&2; exit 1 ;;
esac
done
echo "=== kill visual_servo + camera (D405 exclusive) ==="
_kill_servo
pkill -f 'run_camera_d405' 2>/dev/null || true
pkill -f 'realsense_color_pub' 2>/dev/null || true
pkill -f 'realsense2_camera_node' 2>/dev/null || true
sleep 1
if [[ "${backend}" == "rdk" ]]; then
if ros2 service list 2>/dev/null | grep -q '/set_named_pose'; then
echo "ERROR: RDK 拖拽与 run_real.sh 互斥。请先在那个终端 Ctrl-C 停掉 run_real.sh" >&2
exit 1
fi
else
_need_robot
fi
mkdir -p "${WORKDIR}/images"
if [[ "${append}" == true ]]; then
echo "=== append: keep existing ${TAUGHT} ==="
else
# 重新设点:丢掉上次 taught,避免 merge 混进旧位姿
: > "${TAUGHT}"
fi
echo "=== interactive teach backend=${backend} → ${WORKDIR} ==="
echo " 窗口按键: d=拖拽开 s=拖拽关 Enter=拍照记点 q=结束"
if [[ "${points}" -gt 0 ]] 2>/dev/null; then
echo " 目标 ${points} 个点(采够会有提示,仍可继续多采)"
else
echo " 建议 8–12 个姿态,棋盘占画面约 30–70%"
echo " (也可用 --points N 设定目标点数)"
fi
echo " 结束后: bash ~/ros2/scripts/run_handeye.sh merge"
echo " bash ~/ros2/scripts/run_handeye.sh calibrate"
if [[ "${auto_solve}" == true ]]; then
echo " --auto-solve: 采完会自动 merge,并视后端决定是否直接标定"
fi
TEACH_ARGS=(
-p backend:="${backend}"
-p robot_sn:="${SITE_ROBOT_SN:-Rizon4-063989}"
-p joint_prefix:="${PREFIX}"
-p camera_source:="${CAM}"
-p output_dir:="${WORKDIR}"
-p target_points:="${points}"
)
if [[ -n "${SITE_CAM_SERIAL:-}" ]]; then
TEACH_ARGS+=(-p camera_serial:="${SITE_CAM_SERIAL}")
fi
if [[ -n "${SITE_CAM_WIDTH:-}" ]]; then
TEACH_ARGS+=(-p width:="${SITE_CAM_WIDTH}" -p height:="${SITE_CAM_HEIGHT:-480}" -p fps:="${SITE_CAM_FPS:-30}")
fi
TEACH_ARGS+=(
-p auto_exposure:="$([ "${SITE_CAM_AUTO_EXPOSURE:-1}" = "1" ] && echo true || echo false)"
-p ae_limit_us:="${SITE_CAM_AE_LIMIT_US:-1800}"
-p emitter_enabled:="$([ "${SITE_CAM_EMITTER:-0}" = "1" ] && echo true || echo false)"
-p exposure_us:="${SITE_CAM_EXPOSURE_US:-0}"
-p gain:="${SITE_CAM_GAIN:-0}"
)
ros2 run arm_application handeye_teach --ros-args "${TEACH_ARGS[@]}" || true
# ---- 采点结束 ----
if [[ "${auto_solve}" != true ]]; then
echo
echo "采点结束。接着跑(或下次加 --auto-solve 让它自动接):"
echo " bash ~/ros2/scripts/run_handeye.sh merge --arm ${SITE_ROBOT_ARM:-89}"
echo " bash ~/ros2/scripts/run_handeye.sh calibrate --arm ${SITE_ROBOT_ARM:-89}"
return 0
fi
echo
echo "=== --auto-solve: 自动 merge ==="
if [[ ! -s "${TAUGHT}" ]]; then
echo "ERROR: 没采到任何点(${TAUGHT} 为空),中止。" >&2
return 1
fi
if ! python3 "${ROOT}/scripts/merge_handeye_poses.py" "${TAUGHT}" "${WP}"; then
echo "ERROR: merge 失败,点还在 ${TAUGHT},可手动重跑 merge。" >&2
return 1
fi
echo " 已写入 ${WP}"
# RDK 软拖拽与控制器互斥:软拖拽要 RDK 独占,标定必须在 run_real.sh 的
# 控制器上走点位,两者无法同时存在。这是硬件约束,脚本绕不过去,
# 所以这里如实停下并给出下一步,而不是硬跑一个必然失败的 calibrate。
if [[ "${backend}" == "rdk" ]]; then
echo
echo "⚠️ 当前是 RDK 拖拽模式(--rdk),无法自动接着求解。"
echo " 原因:软拖拽需要 RDK 独占,而标定需要通过 run_real.sh 的控制器走点位,"
echo " 两者不能同时存在。"
echo
echo " 继续两步即可:"
echo " 1) 另开终端: bash ~/ros2/scripts/run_real.sh --arm ${SITE_ROBOT_ARM:-89}"
echo " 2) 回到这里: bash ~/ros2/scripts/run_handeye.sh calibrate --arm ${SITE_ROBOT_ARM:-89}"
echo
echo " 想真正一条命令跑完,请改用示教器 Free Drive 模式:"
echo " bash ~/ros2/scripts/run_handeye.sh teach --ros --auto-solve"
return 0
fi
# --ros(示教器)后端:控制器本来就在跑,可以无缝接上标定。
# 但 teach 阶段把相机杀掉了(USB 独占),标定拍照要用,得先拉回来。
echo
echo "=== 重启相机(teach 阶段杀掉了它,标定拍照需要)==="
nohup bash "${ROOT}/scripts/run_camera_d405.sh" --arm "${SITE_ROBOT_ARM:-89}" \
--no-preview > /tmp/run_handeye_camera.log 2>&1 &
echo " pid=$! log=/tmp/run_handeye_camera.log"
# _need_camera 内部失败时会自己打印原因并 exit,不需要在这里再判一次。
_need_camera
echo
echo "=== --auto-solve: 自动 calibrate ==="
cmd_calibrate
}
case "${CMD}" in
calibrate|calib) cmd_calibrate ;;
waypoints|tour) cmd_waypoints "$@" ;;
teach) cmd_teach "$@" ;;
merge) cmd_merge "$@" ;;
show) cmd_show ;;
-h|--help|help|"") usage; [[ -n "${CMD}" ]] || exit 0; exit 0 ;;
*) echo "Unknown: ${CMD}" >&2; usage >&2; exit 1 ;;
esac