diff --git a/challenge/onsite_competition/README.md b/challenge/onsite_competition/README.md index 76b91ee7..52c14813 100644 --- a/challenge/onsite_competition/README.md +++ b/challenge/onsite_competition/README.md @@ -5,6 +5,11 @@ In this phase, participants’ models will be deployed on **a real robot** to ev --- +## Robot +The robot uses a wheeled chassis. The chassis and camera control code can be found in the provided SDK. The camera’s default resolution is 640×480, and the depth and color images are already aligned. + +The robot is an agilex [RANGER MINI 3.0](https://www.agilex.ai/solutions/1) with RGB camera and a LiDAR sensor. + ## ⚙️ Installation First, install the `InternNav` package: diff --git a/challenge/onsite_competition/captures/rs_depth_mm.png b/challenge/onsite_competition/captures/rs_depth_mm.png new file mode 100644 index 00000000..7c46c358 Binary files /dev/null and b/challenge/onsite_competition/captures/rs_depth_mm.png differ diff --git a/challenge/onsite_competition/captures/rs_depth_vis.png b/challenge/onsite_competition/captures/rs_depth_vis.png new file mode 100644 index 00000000..85a629b8 Binary files /dev/null and b/challenge/onsite_competition/captures/rs_depth_vis.png differ diff --git a/challenge/onsite_competition/onsite_competition_rules_en-US.md b/challenge/onsite_competition/onsite_competition_rules_en-US.md index 13068ce4..ed9c5b98 100644 --- a/challenge/onsite_competition/onsite_competition_rules_en-US.md +++ b/challenge/onsite_competition/onsite_competition_rules_en-US.md @@ -82,8 +82,7 @@ Successfully completing one instruction will add 10 points to the total score, a | Action | Score Impact | |:--|:--| | Successfully reach goal | +10 points | -| Minor scrape with obstacle | –2 points (cannot go below 0 for that instruction) | -| Collision with obstacle | 0 points, navigation terminated for this instruction | +| Minor scrape more than 5 times or Collision with obstacle | 0 points, navigation terminated for this instruction | If there is a trend of continuous collisions, the referee has the right to terminate the robot’s current action. If the collision occurs only once at the end of a forward movement when approaching an obstacle, it may be considered a minor scrape, with the severity of the impact determined by the on-site referee. @@ -92,14 +91,9 @@ If the collision occurs only once at the end of a forward movement when approach The goal is defined as a 2m-radius circular area (no wall crossing). The run is considered successful if the robot stops inside this area. **Ranking Rules (Onsite Competition)**: -- Higher total score ranks higher. -- Tie-breaker: lower total completion time ranks higher. +- The final score is the success rate, calculated as the number of successful instructions divided by the total number of instructions. ## 5.2 Final Results -Final results combine online phase and onsite phase scores using a rank-based point system: -- Points per Rank: -Points = 100 – 5 × (Rank – 1) +Final results combine online phase and onsite phase scores: - Final Score Calculation: -Final Score = (Online Points × 40%) + (Onsite Points × 60%) - -If the final score the same, onsite points break the tie. +Final Score = (Online SR × 40%) + (Onsite SR × 60%) diff --git a/challenge/onsite_competition/onsite_competition_rules_zh-CN.md b/challenge/onsite_competition/onsite_competition_rules_zh-CN.md index 2006999a..acf8fe08 100644 --- a/challenge/onsite_competition/onsite_competition_rules_zh-CN.md +++ b/challenge/onsite_competition/onsite_competition_rules_zh-CN.md @@ -54,30 +54,24 @@ ## 5. 评分细则 ### 5.1 线下赛 -比赛采用加分与扣分相结合的制度。每支队伍初始分为 0 分。每条指令的多次尝试将独立计分，并仅取该指令的最高得分计入总分。 +比赛采用加分制度。每支队伍初始分为 0 分。每条指令的多次尝试将独立计分，并仅取该指令的最高得分计入总分。 **加分规则**： - 成功完成1条指令，总分增加 10 分，并记录该指令的完成时间。（成功条件：在比赛中，每条指令均对应一个目标位置。目标位置定义为半径 2 米的圆形区域（不穿墙），机器人最终停止在该范围即可判定为该条指令导航成功。） **扣分规则**： -- 运行过程中，若机器人轻微擦碰障碍物，该条指令每次扣除 2 分，最低不低于 0 分； -- 若机器人直接撞击障碍物，本次导航将被强制终止，该条指令不计分。 +- 若机器人频繁刮蹭（5次）或直接撞击障碍物，本次导航将被强制终止，该条指令不计分。 | Action | Score | |:--|:--:| | 抵达目标 | +10 | -| 刮蹭 | -2 | -| 连续碰撞 | 本次不计分 | +| 连续刮蹭或撞击 | 本次不计分 | -有连续撞击趋势裁判有权终止当前机器人行为，如果是单次前进行为末尾抵近障碍物，可算一次剐蹭，由现场裁判判断撞击剧烈程度。 +有连续撞击趋势裁判有权终止当前机器人行为，由现场裁判判断撞击剧烈程度。 -**线下排名规则**： -- 总分高者排名靠前； -- 若总分相同，则以成功完成指令所耗总时间更少的队伍排名更高。 +**线下计分规则**： +- 根据得分计算出 Success Rate （SR），得分除以总分得到成功率。 ### 5.2 最终成绩 -最终结果由线上阶段与线下阶段成绩加权计算，采用基于排名的积分制： -- **积分计算方式**：100 - 5×(排名-1)分 -- **最终成绩计算**：最终积分 = 线上积分 × 40% + 线下积分 × 60% - -若最终成绩相同，则以线下积分优先。 +最终结果由线上阶段与线下阶段成绩加权计算： +- **最终成绩计算**：最终积分 = 线上成功率 × 40% + 线下成功率 × 60% diff --git a/challenge/onsite_competition/sdk/main.py b/challenge/onsite_competition/sdk/main.py index 2e9468ef..705329cd 100644 --- a/challenge/onsite_competition/sdk/main.py +++ b/challenge/onsite_competition/sdk/main.py @@ -3,6 +3,7 @@ import sys from real_world_env import RealWorldEnv +from stream import app, start_env from internnav.agent.utils.client import AgentClient from internnav.configs.evaluator.default_config import get_config @@ -21,6 +22,7 @@ def parse_args(): type=str, help='current instruction to follow', ) + parser.add_argument("--tag", type=str, help="tag for the run, saved by the tag name which is team-task-trail") return parser.parse_args() @@ -32,6 +34,7 @@ def load_eval_cfg(config_path, attr_name='eval_cfg'): return getattr(config_module, attr_name) +# TODO add logging for each step, saved by the tag name which is team-task-trail def main(): args = parse_args() print("--- Loading config from:", args.config, "---") @@ -43,16 +46,34 @@ def main(): agent = AgentClient(cfg.agent) # initialize real world env - env = RealWorldEnv(args.instruction) + env = RealWorldEnv() - while True: - # print("get observation...") - # obs contains {rgb, depth, instruction} - obs = env.get_observation() + # start stream + start_env(env) + app.run(host="0.0.0.0", port=8080, threaded=True) - # print("agent step...") - # action is a integer in [0, 3], agent return [{'action': [int], 'ideal_flag': bool}] (same to internvla_n1 agent) - action = agent.step(obs)[0]['action'][0] # only take the first env's action integer + try: + while True: + # print("get observation...") + # obs contains {rgb, depth, instruction} + obs = env.get_observation() + obs["instruction"] = args.instruction - # print("env step...") - env.step(action) + # print("agent step...") + # action is a integer in [0, 3], agent return [{'action': [int], 'ideal_flag': bool}] (same to internvla_n1 agent) + try: + action = agent.step(obs)[0]['action'][0] + print(f"agent step success, action is {action}") + except Exception as e: + print(f"agent step error {e}") + continue + + # print("env step...") + try: + env.step(action) + print("env step success") + except Exception as e: + print(f"env step error {e}") + continue + finally: + env.close() diff --git a/challenge/onsite_competition/sdk/real_world_env.py b/challenge/onsite_competition/sdk/real_world_env.py index 8d13c99c..e004bd64 100644 --- a/challenge/onsite_competition/sdk/real_world_env.py +++ b/challenge/onsite_competition/sdk/real_world_env.py @@ -1,3 +1,6 @@ +import threading +import time + from cam import AlignedRealSense from control import DiscreteRobotController @@ -5,23 +8,49 @@ class RealWorldEnv(Env): - def __init__(self): + def __init__(self, fps: int = 30): self.node = DiscreteRobotController() self.cam = AlignedRealSense() + self.latest_obs = None + self.lock = threading.Lock() + self.stop_flag = threading.Event() + self.fps = fps + + # 启动相机 + self.cam.start() + # 启动采集线程 + self.thread = threading.Thread(target=self._capture_loop, daemon=True) + self.thread.start() + + def _capture_loop(self): + """keep capturing frames""" + interval = 1.0 / self.fps + while not self.stop_flag.is_set(): + t0 = time.time() + try: + obs = self.cam.get_observation(timeout_ms=1000) + with self.lock: + self.latest_obs = obs + except Exception as e: + print("Camera capture failed:", e) + time.sleep(0.05) + dt = time.time() - t0 + if dt < interval: + time.sleep(interval - dt) def get_observation(self): - frame = self.cam.get_observation() - return frame - - def step(self, action): - - ''' - action (int): Discrete action to apply: - - 0: no movement (stand still) - - 1: move forward - - 2: rotate left - - 3: rotate right - ''' + """return most recent frame""" + with self.lock: + return self.latest_obs + + def step(self, action: int): + """ + action: + 0: stand still + 1: move forward + 2: turn left + 3: turn right + """ if action == 0: self.node.stand_still() elif action == 1: @@ -30,3 +59,8 @@ def step(self, action): self.node.turn_left() elif action == 3: self.node.turn_right() + + def close(self): + self.stop_flag.set() + self.thread.join(timeout=1.0) + self.cam.stop() diff --git a/challenge/onsite_competition/sdk/stream.py b/challenge/onsite_competition/sdk/stream.py new file mode 100644 index 00000000..b045cf12 --- /dev/null +++ b/challenge/onsite_competition/sdk/stream.py @@ -0,0 +1,38 @@ +# stream_server.py +import time + +import cv2 +from flask import Flask, Response + +app = Flask(__name__) + +# 由主程序注入 +_env = None + + +def set_env(env): + """set env from main to stream server""" + global _env + _env = env + + +def _mjpeg_generator(jpeg_quality: int = 80): + boundary = b"--frame" + while True: + if _env is None: + time.sleep(0.1) + continue + obs = _env.get_observation() + if obs is None: + time.sleep(0.01) + continue + frame_bgr = obs["rgb"] + ok, jpg = cv2.imencode(".jpg", frame_bgr, [cv2.IMWRITE_JPEG_QUALITY, jpeg_quality]) + if not ok: + continue + yield (boundary + b"\r\n" b"Content-Type: image/jpeg\r\n\r\n" + jpg.tobytes() + b"\r\n") + + +@app.route("/stream") +def stream(): + return Response(_mjpeg_generator(), mimetype="multipart/x-mixed-replace; boundary=frame")