v0.1.7; fix prob of not adjusting resample cmd based on the terrains

UPDATE.md

@@ -1,7 +1,9 @@
 # 20260107
+## v0.1.7
+Fix bug: 修复`cfg.commands.dynamic_resample_commands=False`时, 未根据环境调整指令采样范围的问题
 ## v0.1.6
 1. 加入`go2_rem_cts`, student使用MoE结构, teacher使用普通CTS, 使用非共享权重和全goal输入
-2. 加入`move_down_by_acuumulated_xy_command`选择是否通过累计速度来降低等级
+2. 加入`move_down_by_accumulated_xy_command`选择是否通过累计速度来降低等级
 3. 加入`dynamic_resample_commands`选择是否通过累计速度来动态调整指令采样下限
 # 20260106
 ## v0.1.5

legged_gym/envs/base/legged_robot.py

@@ -16,7 +16,7 @@ from legged_gym import LEGGED_GYM_ROOT_DIR
 from legged_gym.envs.base.base_task import BaseTask
 from legged_gym.utils.math import wrap_to_pi, quat_apply_yaw
 from legged_gym.utils.isaacgym_utils import get_euler_xyz as get_euler_xyz_in_tensor
-from legged_gym.utils.isaacgym_utils import sample_disjoint_intervals
+from legged_gym.utils.isaacgym_utils import sample_disjoint_intervals, sample_single_interval
 from legged_gym.utils.helpers import class_to_dict
 from .legged_robot_config import LeggedRobotCfg
 from legged_gym.utils.terrain import Terrain
@@ -474,19 +474,36 @@ class LeggedRobot(BaseTask):
                 self.commands[env_ids, 2] = (upper - lower) * r + lower
             self.commands_resampling_step[env_ids] = self.cfg.commands.resampling_time / self.dt
         else:
-            self.commands[env_ids, 0] = torch_rand_float(self.command_ranges["lin_vel_x"][0], self.command_ranges["lin_vel_x"][1], (len(env_ids), 1), device=self.device).squeeze(1)
-            self.commands[env_ids, 1] = torch_rand_float(self.command_ranges["lin_vel_y"][0], self.command_ranges["lin_vel_y"][1], (len(env_ids), 1), device=self.device).squeeze(1)
+            self.commands[env_ids, 0] = sample_single_interval(
+                env_ids,
+                self.env_command_ranges["lin_vel_x"][env_ids, 0],
+                self.env_command_ranges["lin_vel_x"][env_ids, 1],
+                self.device
+            )
+            self.commands[env_ids, 1] = sample_single_interval(
+                env_ids,
+                self.env_command_ranges["lin_vel_y"][env_ids, 0],
+                self.env_command_ranges["lin_vel_y"][env_ids, 1],
+                self.device
+            )
             if self.cfg.commands.heading_command:
-                self.commands[env_ids, 3] = torch_rand_float(self.command_ranges["heading"][0], self.command_ranges["heading"][1], (len(env_ids), 1), device=self.device).squeeze(1)
+                self.commands[env_ids, 3] = sample_single_interval(
+                    env_ids,
+                    self.env_command_ranges["heading"][env_ids, 0],
+                    self.env_command_ranges["heading"][env_ids, 1],
+                    self.device
+                )
             else:
-                self.commands[env_ids, 2] = torch_rand_float(self.command_ranges["ang_vel_yaw"][0], self.command_ranges["ang_vel_yaw"][1], (len(env_ids), 1), device=self.device).squeeze(1)
+                self.commands[env_ids, 2] = sample_single_interval(
+                    env_ids,
+                    self.env_command_ranges["ang_vel_yaw"][env_ids, 0],
+                    self.env_command_ranges["ang_vel_yaw"][env_ids, 1],
+                    self.device
+                )
 
             # set small commands to zero
             self.commands[env_ids, :2] *= (torch.norm(self.commands[env_ids, :2], dim=1) > 0.2).unsqueeze(1)
 
-        # set small commands to zero
-        # self.commands[env_ids, :2] *= (torch.norm(self.commands[env_ids, :2], dim=1) > 0.2).unsqueeze(1)
-
         rand_prob = torch.rand(len(env_ids), device=self.device)
         min_prob, max_prob = 0.0, 0.0
         # set limitation lin vel
@@ -1129,7 +1146,7 @@ class LeggedRobot(BaseTask):
         distance = self.max_move_distance[env_ids]
         # robots that walked far enough progress to harder terains
         move_up = distance > self.terrain.env_length / 2
-        if self.cfg.terrain.move_down_by_acuumulated_xy_command:
+        if self.cfg.terrain.move_down_by_accumulated_xy_command:
             move_down = (distance < torch.norm(self.commands_xy_accumulation[env_ids], dim=1) * (self.cfg.commands.resampling_time * (1 - self.zero_command_proba)) * 0.5) * ~move_up
         else:
             # robots that walked less than half of their required distance go to simpler terrains

legged_gym/envs/base/legged_robot_config.py

@@ -38,7 +38,7 @@ class LeggedRobotCfg(BaseConfig):
         terrain_proportions = [0.1, 0.1, 0.1, 0.2, 0.2, 0.1, 0.1, 0.1, 0.0]
         # trimesh only:
         slope_treshold = 0.75 # slopes above this threshold will be corrected to vertical surfaces
-        move_down_by_acuumulated_xy_command = False # move down the terrain curriculum based on accumulated xy command distance instead of absolute distance
+        move_down_by_accumulated_xy_command = False # move down the terrain curriculum based on accumulated xy command distance instead of absolute distance
 
     class commands:
         curriculum = False

legged_gym/envs/go2/go2_config.py

@@ -93,7 +93,7 @@ class GO2Cfg(LeggedRobotCfg):
         # terrain_proportions = [0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0]
         # terrain_proportions = [0.3, 0.3, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1]
         # terrain_proportions = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0]
-        move_down_by_acuumulated_xy_command = True # move down the terrain curriculum based on accumulated xy command distance instead of absolute distance
+        move_down_by_accumulated_xy_command = True # move down the terrain curriculum based on accumulated xy command distance instead of absolute distance
         
     class commands(LeggedRobotCfg.commands):
         curriculum = False

legged_gym/envs/go2/go2_config_vanilla.py

@@ -94,7 +94,7 @@ class GO2Cfg(LeggedRobotCfg):
         # terrain_proportions = [0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0]
         # terrain_proportions = [0.3, 0.3, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1]
         # terrain_proportions = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0]
-        move_down_by_acuumulated_xy_command = False # move down the terrain curriculum based on accumulated xy command distance instead of absolute distance
+        move_down_by_accumulated_xy_command = False # move down the terrain curriculum based on accumulated xy command distance instead of absolute distance
         
     class commands(LeggedRobotCfg.commands):
         curriculum = False

legged_gym/utils/isaacgym_utils.py

@@ -45,3 +45,11 @@ def sample_disjoint_intervals(env_ids, limit_bound, cfg_min, cfg_max, device):
         cfg_max - width_pos + (u - width_neg)
     )
     return samples
+
+def sample_single_interval(env_ids, cfg_min, cfg_max, device):
+    """
+    sample uniform distribution from [cfg_min, cfg_max]
+    """
+    r = torch.rand(len(env_ids), device=device)
+    samples = cfg_min + r * (cfg_max - cfg_min)
+    return samples

setup.py

@@ -2,7 +2,7 @@ from setuptools import find_packages
 from distutils.core import setup
 
 setup(name='go2_rl_gym',
-      version='0.1.5',
+      version='0.1.7',
       author='Wu Tianyang',
       license="MIT",
       packages=find_packages(),