fix: change deps version to fit OpenVLA

.gitignore

@@ -216,3 +216,14 @@ devenv.local.yaml
 
 # pre-commit
 .pre-commit-config.yaml
+
+# Devenv
+.devenv*
+devenv.local.nix
+devenv.local.yaml
+
+# direnv
+.direnv
+
+# pre-commit
+.pre-commit-config.yaml

devenv.lock

@@ -3,10 +3,10 @@
     "devenv": {
       "locked": {
         "dir": "src/modules",
-        "lastModified": 1764449550,
+        "lastModified": 1764927628,
         "owner": "cachix",
         "repo": "devenv",
-        "rev": "dfb58ac03bed07b93f629df55034bc50394d3971",
+        "rev": "247d7027f91368054fb0eefbd755a73d42b66fee",
         "type": "github"
       },
       "original": {
@@ -19,10 +19,10 @@
     "flake-compat": {
       "flake": false,
       "locked": {
-        "lastModified": 1761588595,
+        "lastModified": 1765121682,
         "owner": "edolstra",
         "repo": "flake-compat",
-        "rev": "f387cd2afec9419c8ee37694406ca490c3f34ee5",
+        "rev": "65f23138d8d09a92e30f1e5c87611b23ef451bf3",
         "type": "github"
       },
       "original": {
@@ -34,10 +34,10 @@
     "flake-compat_2": {
       "flake": false,
       "locked": {
-        "lastModified": 1761588595,
+        "lastModified": 1765121682,
         "owner": "edolstra",
         "repo": "flake-compat",
-        "rev": "f387cd2afec9419c8ee37694406ca490c3f34ee5",
+        "rev": "65f23138d8d09a92e30f1e5c87611b23ef451bf3",
         "type": "github"
       },
       "original": {
@@ -55,10 +55,10 @@
         ]
       },
       "locked": {
-        "lastModified": 1763988335,
+        "lastModified": 1765016596,
         "owner": "cachix",
         "repo": "git-hooks.nix",
-        "rev": "50b9238891e388c9fdc6a5c49e49c42533a1b5ce",
+        "rev": "548fc44fca28a5e81c5d6b846e555e6b9c2a5a3c",
         "type": "github"
       },
       "original": {
@@ -89,10 +89,10 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1761313199,
+        "lastModified": 1764580874,
         "owner": "cachix",
         "repo": "devenv-nixpkgs",
-        "rev": "d1c30452ebecfc55185ae6d1c983c09da0c274ff",
+        "rev": "dcf61356c3ab25f1362b4a4428a6d871e84f1d1d",
         "type": "github"
       },
       "original": {
@@ -108,10 +108,10 @@
         "nixpkgs": "nixpkgs_2"
       },
       "locked": {
-        "lastModified": 1763677049,
+        "lastModified": 1765052656,
         "owner": "cachix",
         "repo": "nixpkgs-python",
-        "rev": "159d63dc49a4b12bf85fe0e83011a8b69ba1bcb0",
+        "rev": "04b27dbad2e004cb237db202f21154eea3c4f89f",
         "type": "github"
       },
       "original": {
@@ -122,10 +122,10 @@
     },
     "nixpkgs_2": {
       "locked": {
-        "lastModified": 1764316264,
+        "lastModified": 1764939437,
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "9a7b80b6f82a71ea04270d7ba11b48855681c4b0",
+        "rev": "00d2457e2f608b4be6fe8b470b0a36816324b0ae",
         "type": "github"
       },
       "original": {

devenv.nix

@@ -6,8 +6,8 @@
   ...
 }: {
   packages = with pkgs; [
-    glxinfo
+    mujoco
-    vulkan-tools
+    linuxHeaders
   ];
 
   languages.c.enable = true;
@@ -55,30 +55,33 @@
       );
   };
 
-  env = rec {
+  env = {
-    NIX_LD_LIBRARY_PATH = "$NIX_LD_LIBRARY_PATH:/usr/lib/wsl/lib/";
+    CPATH = "${pkgs.linuxHeaders}/include";
-    LD_LIBRARY_PATH = NIX_LD_LIBRARY_PATH;
+    KERNEL_DIR = "${pkgs.linuxHeaders}/include";
-
-    GLVND = "${pkgs.libglvnd}";
-    MESA = "${pkgs.mesa}";
-    MESAD = MESA;
-
-    GALLIUM_DRIVER = "d3d12";
-    MESA_D3D12_DEFAULT_ADAPTER_NAME = "NVIDIA";
-
-    # LIBGL_ALWAYS_INDIRECT = 0;
-    # __GLX_VENDOR_LIBRARY_NAME = "nvidia";
-    # __NV_PRIME_RENDER_OFFLOAD = 1;
-    # __VK_LAYER_NV_optimus = "NVIDIA_only";
-
-    # PYOPENGL_PLATFORM = "egl";
-    # EGL_PLATFORM = "wayland";
-    # __EGL_VENDOR_LIBRARY_DIRS = "${MESA}/share/glvnd/egl_vendor.d:${GLVND}/share/glvnd/egl_vendor.d";
-    # LIBGL_DRIVERS_PATH = "${MESA}/lib/dri";
   };
 
+  # 创建 .venv 软链接到 .devenv/state/venv
+  scripts.linkVenv.exec = ''
+    if [ -L ".venv" ] || [ ! -e ".venv" ]; then
+        rm -f ".venv" 2>/dev/null || true
+        ln -s ".devenv/state/venv" ".venv"
+        echo "Created symlink: .venv -> .devenv/state/venv"
+    elif [ -e ".venv" ]; then
+        echo "Warning: .venv exists and is not a symlink. Not replacing it."
+        echo "If you want to replace it with a symlink, please remove it manually first."
+    fi
+  '';
+
   # https://devenv.sh/basics/
   enterShell = ''
-    git --version # Use packages
+    export NIX_LD_LIBRARY_PATH="$NIX_LD_LIBRARY_PATH:${lib.makeLibraryPath [ pkgs.linuxHeaders ]}";
+    export LD_LIBRARY_PATH="$NIX_LD_LIBRARY_PATH";
+
+    echo "NIX_LD_LIBRARY_PATH set to $NIX_LD_LIBRARY_PATH"
+    echo "LD_LIBRARY_PATH set to $LD_LIBRARY_PATH"
+    echo "CPATH set to $CPATH"
+    echo "KERNEL_DIR set to $KERNEL_DIR"
+
+    linkVenv
   '';
 }

pyproject.toml

@@ -5,8 +5,15 @@ description = "Add your description here"
 readme = "README.md"
 requires-python = ">=3.12"
 dependencies = [
+    "accelerate>=1.12.0",
+    "mujoco>=3.4.0",
+    "pillow>=12.0.0",
+    "robosuite>=1.5.1",
+    "robosuite-models>=1.0.0",
+    "timm>=0.9.10,<1.0.0",
     "torch>=2.9.1",
     "torchvision>=0.24.1",
+    "transformers==4.40.1",
 ]
 
 [dependency-groups]

src/main.py

@@ -1,126 +1,196 @@
-from typing import cast
+"""
+Simple OpenVLA Demo with Robosuite + MuJoCo
+Uses the official OpenVLA model from HuggingFace
+"""
 
-import genesis as gs
 import numpy as np
+import robosuite as suite
+import torch
+from PIL import Image
+from transformers import AutoModelForVision2Seq, AutoProcessor
 
-########################## init ##########################
+
-gs.init(backend=gs.gs_backend.gpu)
+class OpenVLADemo:
+    """OpenVLA Demo with Robosuite using HuggingFace model"""
+
+    # OpenVLA action names (7-DoF)
+    ACTION_NAMES = [
+        "delta_x",
+        "delta_y",
+        "delta_z",
+        "delta_roll",
+        "delta_pitch",
+        "delta_yaw",
+        "gripper",
+    ]
+
+    def __init__(
+        self,
+        env_name: str = "Lift",
+        robot: str = "Panda",
+        instruction: str = "pick up the red cube",
+        model_name: str = "openvla/openvla-7b",
+    ):
+        self.instruction = instruction
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        print(f"[INFO] Using device: {self.device}")
+
+        # Load OpenVLA model from HuggingFace
+        print(f"[INFO] Loading OpenVLA model: {model_name}")
+        print("[INFO] This may take a while on first run...")
+
+        self.processor = AutoProcessor.from_pretrained(
+            model_name, trust_remote_code=True
+        )
+
+        # Load model with appropriate settings
+        self.model = AutoModelForVision2Seq.from_pretrained(
+            model_name,
+            torch_dtype=torch.bfloat16,
+            low_cpu_mem_usage=True,
+            trust_remote_code=True,
+        ).to(self.device)
+
+        self.model.eval()
+        print("[INFO] OpenVLA model loaded successfully!")
+
+        # Create robosuite environment
+        print(f"[INFO] Creating {env_name} environment with {robot} robot...")
+        self.env = suite.make(
+            env_name=env_name,
+            robots=robot,
+            has_renderer=True,
+            has_offscreen_renderer=True,
+            use_camera_obs=True,
+            camera_names="agentview",
+            camera_heights=224,
+            camera_widths=224,
+            control_freq=20,
+            render_camera="frontview",
+        )
+
+        print(f"[INFO] Environment created! Action dim: {self.env.action_dim}")
+
+    def get_prompt(self) -> str:
+        """Format the prompt for OpenVLA"""
+        return f"In: What action should the robot take to {self.instruction}?\nOut:"
+
+    @torch.no_grad()
+    def get_action(self, obs: dict) -> np.ndarray:
+        """Get action from OpenVLA model"""
+        # Get image from observation and convert to PIL
+        image_array = obs["agentview_image"]
+        # Robosuite returns (H, W, C) uint8 image
+        image = Image.fromarray(image_array)
+
+        # Prepare prompt
+        prompt = self.get_prompt()
+
+        # Process inputs
+        inputs = self.processor(prompt, image).to(self.device, dtype=torch.bfloat16)
+
+        # Predict action using OpenVLA
+        # Note: unnorm_key should match your robot setup
+        # For simulation, we'll use raw normalized actions
+        action = self.model.predict_action(
+            **inputs, do_sample=False, unnorm_key="roboturk"
+        )
+
+        # Action is numpy array of shape (7,)
+        return action
+
+    def print_vla_output(self, step: int, action: np.ndarray, reward: float):
+        """Print detailed VLA output to terminal"""
+        print(f"\n[Step {step:6d}] OpenVLA Output:")
+        print(f'  Instruction: "{self.instruction}"')
+        print("  Actions:")
+        for name, value in zip(self.ACTION_NAMES, action):
+            print(f"    {name:12s}: {value:+.4f}")
+        print(f"  Step Reward: {reward:.4f}")
+
+    def run(self):
+        """Run the demo forever"""
+        print("\n" + "=" * 60)
+        print("OpenVLA Demo Started (Running Forever)")
+        print(f'Instruction: "{self.instruction}"')
+        print("Close the MuJoCo viewer window to quit")
+        print("=" * 60 + "\n")
+
+        episode = 0
+        total_steps = 0
+
+        # Run forever
+        while True:
+            episode += 1
+            print(f"\n{'=' * 60}")
+            print(f"Episode {episode} Started")
+            print("=" * 60)
+
+            obs = self.env.reset()
+            episode_reward = 0.0
+            episode_steps = 0
+
+            while True:
+                # Render environment with MuJoCo viewer
+                self.env.render()
+
+                # Get action from OpenVLA model
+                action = self.get_action(obs)
+
+                # Robosuite expects action of env.action_dim size
+                # OpenVLA outputs 7-DoF, pad or truncate if needed
+                if len(action) < self.env.action_dim:
+                    action = np.concatenate(
+                        [action, np.zeros(self.env.action_dim - len(action))]
+                    )
+                elif len(action) > self.env.action_dim:
+                    action = action[: self.env.action_dim]
+
+                # Step environment
+                obs, reward, done, info = self.env.step(action)
+                episode_reward += reward
+                episode_steps += 1
+                total_steps += 1
+
+                # Print VLA output every 20 steps
+                if episode_steps % 20 == 0:
+                    self.print_vla_output(total_steps, action, reward)
+
+                if done:
+                    print(f"\n[INFO] Episode {episode} done at step {episode_steps}")
+                    print(f"[INFO] Episode Reward: {episode_reward:.3f}")
+                    break
+
+            print(f"\nEpisode {episode} Summary:")
+            print(f"  Steps: {episode_steps}")
+            print(f"  Total Reward: {episode_reward:.3f}")
+            print(f"  Total Steps (all episodes): {total_steps}")
+
+    def close(self):
+        """Clean up"""
+        self.env.close()
+
+
+def main():
+    print("=" * 60)
+    print("OpenVLA Demo with Robosuite + MuJoCo")
+    print("Using OpenVLA-7B from HuggingFace")
+    print("=" * 60)
+
+    demo = OpenVLADemo(
+        env_name="Lift",
+        robot="Panda",
+        instruction="pick up the red cube",
+    )
 
     try:
-    from genesis.engine.entities import RigidEntity
+        demo.run()
-except ImportError:
+    except KeyboardInterrupt:
-    raise ImportError("genesis.engine.entities.RigidEntity is not available")
+        print("\n[INFO] Interrupted by user")
-
+    finally:
-########################## create a scene ##########################
+        demo.close()
-scene = gs.Scene(
+        print("[INFO] Demo closed")
-    viewer_options=gs.options.ViewerOptions(
-        camera_pos=(0, -3.5, 2.5),
-        camera_lookat=(0.0, 0.0, 0.5),
-        camera_fov=30,
-        res=(960, 640),
-        max_FPS=60,
-    ),
-    sim_options=gs.options.SimOptions(
-        dt=0.01,
-    ),
-    show_viewer=True,
-)
-
-########################## entities ##########################
-plane = scene.add_entity(
-    gs.morphs.Plane(),
-)
-franka = scene.add_entity(
-    gs.morphs.MJCF(
-        file="xml/franka_emika_panda/panda.xml",
-    ),
-)
-franka = cast(RigidEntity, franka)
 
 
-########################## build ##########################
+if __name__ == "__main__":
-scene.build()
+    main()
-
-jnt_names = [
-    "joint1",
-    "joint2",
-    "joint3",
-    "joint4",
-    "joint5",
-    "joint6",
-    "joint7",
-    "finger_joint1",
-    "finger_joint2",
-]
-dofs_idx = [franka.get_joint(name).dof_idx_local for name in jnt_names]
-
-############ Optional: set control gains ############
-# set positional gains
-franka.set_dofs_kp(
-    kp=np.array([4500, 4500, 3500, 3500, 2000, 2000, 2000, 100, 100]),
-    dofs_idx_local=dofs_idx,
-)
-# set velocity gains
-franka.set_dofs_kv(
-    kv=np.array([450, 450, 350, 350, 200, 200, 200, 10, 10]),
-    dofs_idx_local=dofs_idx,
-)
-# set force range for safety
-franka.set_dofs_force_range(
-    lower=np.array([-87, -87, -87, -87, -12, -12, -12, -100, -100]),
-    upper=np.array([87, 87, 87, 87, 12, 12, 12, 100, 100]),
-    dofs_idx_local=dofs_idx,
-)
-# Hard reset
-for i in range(150):
-    if i < 50:
-        franka.set_dofs_position(np.array([1, 1, 0, 0, 0, 0, 0, 0.04, 0.04]), dofs_idx)
-    elif i < 100:
-        franka.set_dofs_position(
-            np.array([-1, 0.8, 1, -2, 1, 0.5, -0.5, 0.04, 0.04]), dofs_idx
-        )
-    else:
-        franka.set_dofs_position(np.array([0, 0, 0, 0, 0, 0, 0, 0, 0]), dofs_idx)
-
-    scene.step()
-
-# PD control
-for i in range(1250):
-    if i == 0:
-        franka.control_dofs_position(
-            np.array([1, 1, 0, 0, 0, 0, 0, 0.04, 0.04]),
-            dofs_idx,
-        )
-    elif i == 250:
-        franka.control_dofs_position(
-            np.array([-1, 0.8, 1, -2, 1, 0.5, -0.5, 0.04, 0.04]),
-            dofs_idx,
-        )
-    elif i == 500:
-        franka.control_dofs_position(
-            np.array([0, 0, 0, 0, 0, 0, 0, 0, 0]),
-            dofs_idx,
-        )
-    elif i == 750:
-        # control first dof with velocity, and the rest with position
-        franka.control_dofs_position(
-            np.array([0, 0, 0, 0, 0, 0, 0, 0, 0])[1:],
-            dofs_idx[1:],
-        )
-        franka.control_dofs_velocity(
-            np.array([1.0, 0, 0, 0, 0, 0, 0, 0, 0])[:1],
-            dofs_idx[:1],
-        )
-    elif i == 1000:
-        franka.control_dofs_force(
-            np.array([0, 0, 0, 0, 0, 0, 0, 0, 0]),
-            dofs_idx,
-        )
-    # This is the control force computed based on the given control command
-    # If using force control, it's the same as the given control command
-    print("control force:", franka.get_dofs_control_force(dofs_idx))
-
-    # This is the actual force experienced by the dof
-    print("internal force:", franka.get_dofs_force(dofs_idx))
-
-    scene.step()