Add Go1 training and stair fine-tuning

legged_gym/envs/__init__.py

@@ -2,6 +2,8 @@ from legged_gym import LEGGED_GYM_ROOT_DIR, LEGGED_GYM_ENVS_DIR
 
 from legged_gym.envs.go2.go2_env import Go2Robot
 from legged_gym.envs.go2.go2_config import GO2Cfg, GO2CfgPPO, GO2CfgCTS, GO2CfgMoECTS, GO2CfgMoENGCTS, GO2CfgMCPCTS, GO2CfgACMoECTS, GO2CfgDualMoECTS
+from legged_gym.envs.go1.go1_env import Go1Robot
+from legged_gym.envs.go1.go1_config import GO1Cfg, GO1CfgMoECTS, GO1StairsCfg, GO1StairsCfgMoECTS
 from .base.legged_robot import LeggedRobot
 
 from legged_gym.utils.task_registry import task_registry
@@ -13,3 +15,5 @@ task_registry.register("go2_moe_ng_cts", Go2Robot, GO2Cfg(), GO2CfgMoENGCTS())
 task_registry.register("go2_mcp_cts", Go2Robot, GO2Cfg(), GO2CfgMCPCTS())
 task_registry.register("go2_ac_moe_cts", Go2Robot, GO2Cfg(), GO2CfgACMoECTS())
 task_registry.register("go2_dual_moe_cts", Go2Robot, GO2Cfg(), GO2CfgDualMoECTS())
+task_registry.register("go1_moe_cts", Go1Robot, GO1Cfg(), GO1CfgMoECTS())
+task_registry.register("go1_moe_cts_stairs", Go1Robot, GO1StairsCfg(), GO1StairsCfgMoECTS())

legged_gym/envs/go1/go1_config.py

@@ -0,0 +1,58 @@
+from legged_gym.envs.go2.go2_config import GO2Cfg, GO2CfgMoECTS
+
+
+class GO1Cfg(GO2Cfg):
+    class init_state(GO2Cfg.init_state):
+        pos = [0.0, 0.0, 0.34]
+
+    class control(GO2Cfg.control):
+        stiffness = {'joint': 28.0}
+        damping = {'joint': 0.7}
+        action_scale = 0.25
+        decimation = 4
+
+    class asset(GO2Cfg.asset):
+        file = '{LEGGED_GYM_ROOT_DIR}/resources/robots/go1/urdf/go1.urdf'
+        name = 'go1'
+        self_collisions = 1
+
+    class rewards(GO2Cfg.rewards):
+        base_height_target = 0.30
+
+
+class GO1CfgMoECTS(GO2CfgMoECTS):
+    class runner(GO2CfgMoECTS.runner):
+        experiment_name = 'go1_moe_cts'
+        run_name = 'from_scratch_5000'
+        max_iterations = 5000
+        save_interval = 500
+        resume = False
+        load_run = -1
+        checkpoint = -1
+        resume_path = None
+
+
+class GO1StairsCfg(GO1Cfg):
+    class terrain(GO1Cfg.terrain):
+        # wave, slope, rough slope, stairs up, stairs down, obstacles,
+        # stepping stones, gap, flat
+        terrain_proportions = [0.03, 0.06, 0.03, 0.50, 0.20, 0.08, 0.0, 0.0, 0.10]
+
+    class commands(GO1Cfg.commands):
+        # The 10k checkpoint has only seen +/-0.5 m/s because the original
+        # curriculum does not widen until iteration 20000. Cover the 0.8 m/s
+        # RoboGauge stair command immediately during this fine-tune.
+        command_range_curriculum = []
+
+        class ranges(GO1Cfg.commands.ranges):
+            lin_vel_x = [-1.0, 1.0]
+            lin_vel_y = [-0.5, 0.5]
+            ang_vel_yaw = [-1.0, 1.0]
+
+
+class GO1StairsCfgMoECTS(GO1CfgMoECTS):
+    class runner(GO1CfgMoECTS.runner):
+        experiment_name = 'go1_moe_cts'
+        run_name = 'stairs_finetune_10000_to_15000'
+        max_iterations = 5000
+        save_interval = 500

legged_gym/envs/go1/go1_env.py

@@ -0,0 +1,34 @@
+from legged_gym.envs.go2.go2_env import Go2Robot
+import torch
+
+
+class Go1Robot(Go2Robot):
+    """Go1 asset with an explicit policy-to-asset joint permutation.
+
+    Isaac Gym loads this asset as [FL, FR, RL, RR]. The permutation below
+    presents observations to the policy as [FR, FL, RR, RL] and maps policy
+    actions back to the Isaac Gym asset order. Existing Go1 checkpoints depend
+    on this policy order.
+    """
+
+    # Policy order [FR, FL, RR, RL] -> asset order [FL, FR, RL, RR].
+    _POLICY_TO_ASSET = [3, 4, 5, 0, 1, 2, 9, 10, 11, 6, 7, 8]
+
+    def step(self, actions):
+        asset_actions = torch.zeros_like(actions)
+        asset_actions[:, self._POLICY_TO_ASSET] = actions
+        return super().step(asset_actions)
+
+    def compute_observations(self):
+        super().compute_observations()
+
+        # Go2 policy observation: angular velocity, gravity, commands,
+        # joint-position error, joint velocity, previous action.
+        for start in (9, 21, 33):
+            self.obs_buf[:, start:start + 12] = self.obs_buf[:, start:start + 12][:, self._POLICY_TO_ASSET]
+
+        # Keep the privileged teacher input consistent as well. Its first
+        # three fields are base linear velocity, so joint fields start at 12.
+        if self.privileged_obs_buf is not None:
+            for start in (12, 24, 36, 52, 64):
+                self.privileged_obs_buf[:, start:start + 12] = self.privileged_obs_buf[:, start:start + 12][:, self._POLICY_TO_ASSET]