[Eval, Browser] Refactor Browser Env so it works with EventStreamRuntime for Browsing Evaluation (#3235)

* refactor browser env so it works with eventstream runtime for eval

* fix browsergym environment

opendevin/core/config.py

@@ -154,6 +154,9 @@ class SandboxConfig(metaclass=Singleton):
         initialize_plugins: Whether to initialize plugins.
         update_source_code: Whether to update the source code in the EventStreamRuntime.
             Used for development of EventStreamRuntime.
+        browsergym_eval_env: The BrowserGym environment to use for evaluation.
+            Default is None for general purpose browsing. Check evaluation/miniwob and evaluation/webarena for examples.
+
     """
 
     box_type: str = 'ssh'
@@ -170,6 +173,7 @@ class SandboxConfig(metaclass=Singleton):
     use_host_network: bool = False
     initialize_plugins: bool = True
     update_source_code: bool = False
+    browsergym_eval_env: str | None = None
 
     def defaults_to_dict(self) -> dict:
         """Serialize fields to a dict for the frontend, including type hints, defaults, and whether it's optional."""

opendevin/events/observation/browse.py

@@ -11,7 +11,6 @@ class BrowserOutputObservation(Observation):
 
     url: str
     screenshot: str = field(repr=False)  # don't show in repr
-    status_code: int = 200
     error: bool = False
     observation: str = ObservationType.BROWSE
     # do not include in the memory
@@ -34,12 +33,12 @@ class BrowserOutputObservation(Observation):
         return (
             '**BrowserOutputObservation**\n'
             f'URL: {self.url}\n'
-            f'Status code: {self.status_code}\n'
             f'Error: {self.error}\n'
             f'Open pages: {self.open_pages_urls}\n'
             f'Active page index: {self.active_page_index}\n'
             f'Last browser action: {self.last_browser_action}\n'
             f'Last browser action error: {self.last_browser_action_error}\n'
             f'Focused element bid: {self.focused_element_bid}\n'
+            f'axTree: {self.axtree_object}\n'
             f'CONTENT: {self.content}\n'
         )

opendevin/runtime/browser/browser_env.py

@@ -3,7 +3,6 @@ import base64
 import io
 import json
 import multiprocessing
-import os
 import time
 import uuid
 
@@ -18,41 +17,24 @@ from PIL import Image
 from opendevin.core.exceptions import BrowserInitException
 from opendevin.core.logger import opendevin_logger as logger
 
+BROWSER_EVAL_GET_GOAL_ACTION = 'GET_EVAL_GOAL'
+BROWSER_EVAL_GET_REWARDS_ACTION = 'GET_EVAL_REWARDS'
+
 
 class BrowserEnv:
-    def __init__(
-        self,
-        browsergym_eval: str = '',
-        browsergym_eval_save_dir: str = '',
-    ):
+    def __init__(self, browsergym_eval_env: str | None = None):
         self.html_text_converter = self.get_html_text_converter()
         self.eval_mode = False
         self.eval_dir = ''
-        # EVAL only: browsergym_eval and browsergym_eval_save_dir must be provided for evaluation
-        self.browsergym_eval = browsergym_eval
-        self.browsergym_eval_save_dir = browsergym_eval_save_dir
-        if self.browsergym_eval:
-            assert (
-                self.browsergym_eval_save_dir
-            ), 'browsergym_eval_save_dir must be provided for evaluation.'
-            self.eval_mode = True
-            self.eval_dir = os.path.join(
-                self.browsergym_eval_save_dir, self.browsergym_eval.split('/')[1]
-            )
-            os.makedirs(self.eval_dir, exist_ok=True)
+
+        # EVAL only: browsergym_eval_env must be provided for evaluation
+        self.browsergym_eval_env = browsergym_eval_env
+        self.eval_mode = bool(browsergym_eval_env)
+
         # Initialize browser environment process
         multiprocessing.set_start_method('spawn', force=True)
         self.browser_side, self.agent_side = multiprocessing.Pipe()
 
-        try:
-            self.original_cwd = os.getcwd()
-        except FileNotFoundError:
-            logger.warning(
-                'Current working directory does not exist. Using /tmp as fallback.'
-            )
-            self.original_cwd = '/tmp'
-            os.chdir('/tmp')
-
         self.init_browser()
         atexit.register(self.close)
 
@@ -74,17 +56,6 @@ class BrowserEnv:
     )
     def init_browser(self):
         logger.info('Starting browser env...')
-
-        # Ensure we're in a valid directory before starting the process
-        try:
-            os.chdir(self.original_cwd)
-            logger.debug(f'Changed back to original directory: {self.original_cwd}')
-        except Exception as e:
-            logger.error(f'Failed to change to original directory: {e}')
-            # If we can't change to the original directory, try to use a known valid directory
-            os.chdir('/tmp')
-            logger.debug('Changed to /tmp directory as fallback')
-
         try:
             self.process = multiprocessing.Process(target=self.browser_process)
             self.process.start()
@@ -98,8 +69,17 @@ class BrowserEnv:
 
     def browser_process(self):
         if self.eval_mode:
-            logger.info('Creating browser env for evaluation purpose.')
-            env = gym.make(self.browsergym_eval)
+            assert self.browsergym_eval_env is not None
+            logger.info('Initializing browser env for web browsing evaluation.')
+            if 'webarena' in self.browsergym_eval_env:
+                import browsergym.webarena  # noqa F401 register webarena tasks as gym environments
+            elif 'miniwob' in self.browsergym_eval_env:
+                import browsergym.miniwob  # noqa F401 register miniwob tasks as gym environments
+            else:
+                raise ValueError(
+                    f'Unsupported browsergym eval env: {self.browsergym_eval_env}'
+                )
+            env = gym.make(self.browsergym_eval_env)
         else:
             env = gym.make(
                 'browsergym/openended',
@@ -108,20 +88,22 @@ class BrowserEnv:
                 headless=True,
                 disable_env_checker=True,
             )
+
         obs, info = env.reset()
-        # EVAL only: save the goal into file for evaluation
+
+        # EVAL ONLY: save the goal into file for evaluation
+        self.eval_goal = None
+        self.eval_rewards: list[float] = []
         if self.eval_mode:
-            rewards = []  # store rewards if in eval mode
-            logger.info(obs['goal'])
-            with open(
-                os.path.join(self.eval_dir, 'goal.txt'), 'w', encoding='utf-8'
-            ) as f:
-                f.write(obs['goal'])
+            logger.info(f"Browsing goal: {obs['goal']}")
+            self.eval_goal = obs['goal']
+
         logger.info('Browser env started.')
         while True:
             try:
                 if self.browser_side.poll(timeout=0.01):
                     unique_request_id, action_data = self.browser_side.recv()
+
                     # shutdown the browser environment
                     if unique_request_id == 'SHUTDOWN':
                         logger.info('SHUTDOWN recv, shutting down browser env...')
@@ -130,17 +112,29 @@ class BrowserEnv:
                     elif unique_request_id == 'IS_ALIVE':
                         self.browser_side.send(('ALIVE', None))
                         continue
+
+                    # EVAL ONLY: Get evaluation info
+                    if action_data['action'] == BROWSER_EVAL_GET_GOAL_ACTION:
+                        self.browser_side.send(
+                            (unique_request_id, {'text_content': self.eval_goal})
+                        )
+                        continue
+                    elif action_data['action'] == BROWSER_EVAL_GET_REWARDS_ACTION:
+                        self.browser_side.send(
+                            (
+                                unique_request_id,
+                                {'text_content': json.dumps(self.eval_rewards)},
+                            )
+                        )
+                        continue
+
                     action = action_data['action']
                     obs, reward, terminated, truncated, info = env.step(action)
-                    # EVAL only: save the rewards into file for evaluation
+
+                    # EVAL ONLY: Save the rewards into file for evaluation
                     if self.eval_mode:
-                        rewards.append(reward)
-                        with open(
-                            os.path.join(self.eval_dir, 'rewards.json'),
-                            'w',
-                            encoding='utf-8',
-                        ) as f:
-                            f.write(json.dumps(rewards))
+                        self.eval_rewards.append(reward)
+
                     # add text content of the page
                     html_str = flatten_dom_to_str(obs['dom_object'])
                     obs['text_content'] = self.html_text_converter.handle(html_str)
@@ -158,6 +152,7 @@ class BrowserEnv:
                 return
 
     def step(self, action_str: str, timeout: float = 30) -> dict:
+        """Execute an action in the browser environment and return the observation."""
         unique_request_id = str(uuid.uuid4())
         self.agent_side.send((unique_request_id, {'action': action_str}))
         start_time = time.time()

opendevin/runtime/browser/utils.py

@@ -32,21 +32,23 @@ async def browse(
         obs = browser.step(action_str)
         return BrowserOutputObservation(
             content=obs['text_content'],  # text content of the page
-            open_pages_urls=obs['open_pages_urls'],  # list of open pages
-            active_page_index=obs['active_page_index'],  # index of the active page
-            dom_object=obs['dom_object'],  # DOM object
-            axtree_object=obs['axtree_object'],  # accessibility tree object
-            extra_element_properties=obs[
-                'extra_element_properties'
-            ],  # extra element properties
-            last_browser_action=obs['last_action'],  # last browser env action performed
-            focused_element_bid=obs['focused_element_bid'],  # focused element bid
-            screenshot=obs['screenshot'],  # base64-encoded screenshot, png
-            url=obs['url'],  # URL of the page
-            error=True if obs['last_action_error'] else False,  # error flag
-            last_browser_action_error=obs[
-                'last_action_error'
-            ],  # last browser env action error
+            url=obs.get('url', ''),  # URL of the page
+            screenshot=obs.get('screenshot', None),  # base64-encoded screenshot, png
+            open_pages_urls=obs.get('open_pages_urls', []),  # list of open pages
+            active_page_index=obs.get(
+                'active_page_index', -1
+            ),  # index of the active page
+            dom_object=obs.get('dom_object', {}),  # DOM object
+            axtree_object=obs.get('axtree_object', {}),  # accessibility tree object
+            extra_element_properties=obs.get('extra_element_properties', {}),
+            focused_element_bid=obs.get(
+                'focused_element_bid', None
+            ),  # focused element bid
+            last_browser_action=obs.get(
+                'last_action', ''
+            ),  # last browser env action performed
+            last_browser_action_error=obs.get('last_action_error', ''),
+            error=True if obs.get('last_action_error', '') else False,  # error flag
         )
     except Exception as e:
         return BrowserOutputObservation(

opendevin/runtime/client/client.py

@@ -64,7 +64,12 @@ class RuntimeClient:
     """
 
     def __init__(
-        self, plugins_to_load: list[Plugin], work_dir: str, username: str, user_id: int
+        self,
+        plugins_to_load: list[Plugin],
+        work_dir: str,
+        username: str,
+        user_id: int,
+        browsergym_eval_env: str | None,
     ) -> None:
         self.plugins_to_load = plugins_to_load
         self.username = username
@@ -74,7 +79,7 @@ class RuntimeClient:
         self._init_bash_shell(self.pwd, self.username)
         self.lock = asyncio.Lock()
         self.plugins: dict[str, Plugin] = {}
-        self.browser = BrowserEnv()
+        self.browser = BrowserEnv(browsergym_eval_env)
 
     async def ainit(self):
         for plugin in self.plugins_to_load:
@@ -362,6 +367,12 @@ if __name__ == '__main__':
         '--username', type=str, help='User to run as', default='opendevin'
     )
     parser.add_argument('--user-id', type=int, help='User ID to run as', default=1000)
+    parser.add_argument(
+        '--browsergym-eval-env',
+        type=str,
+        help='BrowserGym environment used for browser evaluation',
+        default=None,
+    )
     # example: python client.py 8000 --working-dir /workspace --plugins JupyterRequirement
     args = parser.parse_args()
 
@@ -382,6 +393,7 @@ if __name__ == '__main__':
             work_dir=args.working_dir,
             username=args.username,
             user_id=args.user_id,
+            browsergym_eval_env=args.browsergym_eval_env,
         )
         await client.ainit()
         yield

opendevin/runtime/client/runtime.py

@@ -147,6 +147,12 @@ class EventStreamRuntime(Runtime):
 
             logger.info(f'run_as_devin: `{self.config.run_as_devin}`')
 
+            if self.config.sandbox.browsergym_eval_env is not None:
+                browsergym_arg = (
+                    f'--browsergym-eval-env {self.config.sandbox.browsergym_eval_env}'
+                )
+            else:
+                browsergym_arg = ''
             container = self.docker_client.containers.run(
                 self.container_image,
                 command=(
@@ -156,7 +162,8 @@ class EventStreamRuntime(Runtime):
                     f'--working-dir {sandbox_workspace_dir} '
                     f'{plugin_arg}'
                     f'--username {"opendevin" if self.config.run_as_devin else "root"} '
-                    f'--user-id {self.config.sandbox.user_id}'
+                    f'--user-id {self.config.sandbox.user_id} '
+                    f'{browsergym_arg}'
                 ),
                 network_mode=network_mode,
                 ports=port_mapping,

tests/unit/test_runtime.py

@@ -1,6 +1,7 @@
 """Test the EventStreamRuntime, which connects to the RuntimeClient running in the sandbox."""
 
 import asyncio
+import json
 import os
 import tempfile
 import time
@@ -13,6 +14,7 @@ from opendevin.core.config import AppConfig, SandboxConfig, load_from_env
 from opendevin.core.logger import opendevin_logger as logger
 from opendevin.events import EventStream
 from opendevin.events.action import (
+    BrowseInteractiveAction,
     BrowseURLAction,
     CmdRunAction,
     FileReadAction,
@@ -29,6 +31,7 @@ from opendevin.events.observation import (
 )
 from opendevin.runtime.client.runtime import EventStreamRuntime
 from opendevin.runtime.plugins import AgentSkillsRequirement, JupyterRequirement
+from opendevin.runtime.runtime import Runtime
 from opendevin.runtime.server.runtime import ServerRuntime
 from opendevin.storage import get_file_store
 
@@ -95,7 +98,8 @@ async def _load_runtime(
     run_as_devin: bool = True,
     enable_auto_lint: bool = False,
     container_image: str | None = None,
-):
+    browsergym_eval_env: str | None = None,
+) -> Runtime:
     sid = 'test'
     cli_session = 'main_test'
     # AgentSkills need to be initialized **before** Jupyter
@@ -104,7 +108,10 @@ async def _load_runtime(
     config = AppConfig(
         workspace_base=temp_dir,
         workspace_mount_path=temp_dir,
-        sandbox=SandboxConfig(use_host_network=True),
+        sandbox=SandboxConfig(
+            use_host_network=True,
+            browsergym_eval_env=browsergym_eval_env,
+        ),
     )
     load_from_env(config, os.environ)
     config.run_as_devin = run_as_devin
@@ -120,7 +127,9 @@ async def _load_runtime(
         # NOTE: we will use the default container image specified in the config.sandbox
         # if it is an official od_runtime image.
         cur_container_image = config.sandbox.container_image
-        if 'od_runtime' not in cur_container_image:
+        if 'od_runtime' not in cur_container_image and cur_container_image not in {
+            'xingyaoww/od-eval-miniwob:v1.0'
+        }:  # a special exception list
             cur_container_image = 'ubuntu:22.04'
             logger.warning(
                 f'`{config.sandbox.container_image}` is not an od_runtime image. Will use `{cur_container_image}` as the container image for testing.'
@@ -387,7 +396,6 @@ async def test_simple_browse(temp_dir, box_class, run_as_devin):
 
     assert isinstance(obs, BrowserOutputObservation)
     assert 'http://localhost:8000' in obs.url
-    assert obs.status_code == 200
     assert not obs.error
     assert obs.open_pages_urls == ['http://localhost:8000/']
     assert obs.active_page_index == 0
@@ -407,6 +415,53 @@ async def test_simple_browse(temp_dir, box_class, run_as_devin):
     await asyncio.sleep(1)
 
 
+@pytest.mark.asyncio
+async def test_browsergym_eval_env(temp_dir):
+    runtime = await _load_runtime(
+        temp_dir,
+        # only supported in event stream runtime
+        box_class=EventStreamRuntime,
+        run_as_devin=False,  # need root permission to access file
+        container_image='xingyaoww/od-eval-miniwob:v1.0',
+        browsergym_eval_env='browsergym/miniwob.choose-list',
+    )
+    from opendevin.runtime.browser.browser_env import (
+        BROWSER_EVAL_GET_GOAL_ACTION,
+        BROWSER_EVAL_GET_REWARDS_ACTION,
+    )
+
+    # Test browse
+    action = BrowseInteractiveAction(browser_actions=BROWSER_EVAL_GET_GOAL_ACTION)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = await runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+
+    assert isinstance(obs, BrowserOutputObservation)
+    assert not obs.error
+    assert 'Select' in obs.content
+    assert 'from the list and click Submit' in obs.content
+
+    # Make sure the browser can produce observation in eva[l
+    action = BrowseInteractiveAction(browser_actions='noop()')
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = await runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert (
+        obs.url.strip()
+        == 'file:///miniwob-plusplus/miniwob/html/miniwob/choose-list.html'
+    )
+
+    # Make sure the rewards are working
+    action = BrowseInteractiveAction(browser_actions=BROWSER_EVAL_GET_REWARDS_ACTION)
+    logger.info(action, extra={'msg_type': 'ACTION'})
+    obs = await runtime.run_action(action)
+    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
+    assert json.loads(obs.content) == [0.0]
+
+    await runtime.close()
+    await asyncio.sleep(1)
+
+
 @pytest.mark.asyncio
 async def test_single_multiline_command(temp_dir, box_class):
     runtime = await _load_runtime(temp_dir, box_class)