update rules; update rgb obs

Author: kew6688

challenge/onsite_competition/onsite_competition_rules_en-US.md

@@ -38,13 +38,12 @@ The challenge centers on vision-language fusion for cross-room end-to-end naviga
 - A standardized debugging time and environment will be provided onsite. Teams may use model weights different from the online stage and make environment-specific configuration adjustments, but modifications to the core algorithm logic are strictly prohibited.
 
 ### 4.2 Procedure
-Each team will receive 10 fixed-sequence instructions. Start from the first one.
+Each team will receive 10 instructions. Start from any instruction.
 For each instruction:
 - Move the robot to the given starting position.
 - Provide the instruction to the robot and raise your hand to signal the start.
-- If execution fails, teams may retry (up to 3 attempts) or skip the instruction.
-- After 3 failed attempts (timeout, collision, human intervention, etc.), the instruction must be skipped.
-- Skipped instructions cannot be retried later.
+- If execution fails, teams may retry or skip the instruction.
+- Instruction can be repeated if failed (timeout, collision, human intervention, etc.).
 - Before each attempt, the algorithm state must be reset and previous observations cleared.
 
 ### 4.3 Time Limit
@@ -82,7 +81,7 @@ Successfully completing one instruction will be one success, and the completion
 | Action | Score Impact |
 |:--|:--|
 | Successfully reach goal | success |
-| Minor scrape more than 5 times or Collision with obstacle | fail |
+| Minor scrape or Collision with obstacle | fail |
 
 If there is a trend of continuous collisions, the referee has the right to terminate the robot’s current action, with the severity of the impact determined by the on-site referee.
 

challenge/onsite_competition/onsite_competition_rules_zh-CN.md

@@ -32,11 +32,10 @@
 - 比赛现场将提供统一的调试时间和环境，允许使用与线上赛不同的模型权重，允许进行环境适配性配置修改，但禁止修改核心算法逻辑。
 
 ### 4.2 比赛过程
-每支队伍开始挑战时将会提供十条固定顺序的指令，从第一条指令开始。
+每支队伍开始挑战时将会提供十条导航指令，选手可选择指令进行。
   - 移动机器人到达该指令的起始位置后，将指令输入给机器人后举手示意作为开始挑战的信号。
-  - 指令执行过程中若中途失败，可选择重试，也可以选择跳过该指令，顺序执行下一条未执行指令。
-  - 每条指令可重试3次，若3次失败，则必须跳过该指令。（失败指选手选择重新开始，或其他条件，如超时，冲撞障碍物，人为干预等）
-  - 已经跳过的指令不能重复执行。
+  - 指令执行过程中若中途失败，可选择重试，也可以选择跳过该指令。
+  - 失败时可选择重新开始（如超时，冲撞障碍物，人为干预等）
   - 每次测试前需重置算法状态并清除所有前序执行过程中获取的观测缓存。
 
 ### 4.3 时间限制
@@ -60,7 +59,7 @@
 - 成功完成1条指令，算成功一次，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**：
-- 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，该条指令失败。
+- 若机器人连续刮蹭或直接撞击障碍物，本次导航将被强制终止，该条指令失败。
 
 | Action | Score |
 |:--|:--:|

challenge/onsite_competition/sdk/cam.py

@@ -99,7 +99,8 @@ def get_observation(self, timeout_ms: int = 1000) -> Dict:
         if not color or not depth:
             raise RuntimeError("can not align color/depth frame")
 
-        rgb = np.asanyarray(color.get_data())  # HxWx3, uint8 (BGR)
+        bgr = np.asanyarray(color.get_data())  # HxWx3, uint8 (BGR)
+        rgb = bgr[..., ::-1]  # HxWx3, uint8 (convert to RGB)
         depth_raw = np.asanyarray(depth.get_data())  # HxW, uint16
         if depth_raw.shape != rgb.shape[:2]:
             # Extreme fallback (theoretically should be consistent after alignment).