refined prompts (6e707706) · Commits · 张泽凯 / VAGEN

vagen/env/frozenlake/env.py

+2 −2

Original line number	Diff line number	Diff line
		@@ -197,9 +197,9 @@ class FrozenLakeEnv(BaseEnv):
		str: System prompt string with environment description and instructions
		"""
		if self.config.render_mode == 'vision':
		return system_prompt_vision.format(max_actions_per_step=self.config.max_actions_per_step)
		return system_prompt_vision.format(max_actions_per_step=self.config.max_actions_per_step,action_sep=self.config.action_sep)
		else:
		return system_prompt_text.format(max_actions_per_step=self.config.max_actions_per_step)
		return system_prompt_text.format(max_actions_per_step=self.config.max_actions_per_step,action_sep=self.config.action_sep)

		def compute_reward(self):
		"""

vagen/env/frozenlake/prompt.py

+4 −2

Original line number	Diff line number	Diff line
		@@ -10,7 +10,8 @@ Rules:
		1. Avoid falling into holes (O).
		2. Frozen tiles are slippery, you may move perpendicular to your intended direction.

		Actions you can take: Left, Down, Right, Up. You can take up to {max_actions_per_step} action(s) at a time.
		Actions you can take: Left, Down, Right, Up.
		You can take up to {max_actions_per_step} action(s) at a time, separated by {action_sep}.
		Left: move left to the cell to the left.
		Down: move down to the cell below.
		Right: move right to the cell to the right.
		@@ -37,7 +38,8 @@ Rules:
		1. Avoid falling into holes.
		2. Frozen tiles are slippery, you may move perpendicular to your intended direction.

		Actions you can take: Left, Down, Right, Up. You can take up to {max_actions_per_step} action(s) at a time.
		Actions you can take: Left, Down, Right, Up.
		You can take up to {max_actions_per_step} action(s) at a time, separated by {action_sep}.

		Rewards:
		Fall into hole: 0

vagen/env/navigation/env.py

+4 −2

Original line number	Diff line number	Diff line
		@@ -349,11 +349,13 @@ class NavigationEnv(BaseEnv):
		"""
		if self.config.visual_env:
		return system_prompt_vision.format(
		max_actions_per_step=self.config.max_actions_per_step
		max_actions_per_step=self.config.max_actions_per_step,
		action_sep=self.config.action_sep
		)
		else:
		return system_prompt_text.format(
		max_actions_per_step=self.config.max_actions_per_step
		max_actions_per_step=self.config.max_actions_per_step,
		action_sep=self.config.action_sep
		)

		def compute_reward(self):

vagen/env/navigation/prompt.py

+2 −3

Original line number	Diff line number	Diff line
		@@ -4,7 +4,6 @@ Navigation Guide
		Goal: Achieve the human instruction

		Actions you can take: moveahead, moveback, moveright, moveleft, rotateright, rotateleft, lookup, lookdown.
		You can take up to {max_actions_per_step} action(s) at a time.

		moveahead: Move forward by 0.25 meter
		moveback: Move backward by 0.25 meter
		@@ -21,10 +20,10 @@ Achieve the human instruction: +10.0

		Please think step by step and provide the actions you want to take.

		You can provide a list of actions, separated by commas.
		You can take up to {max_actions_per_step} action(s) at a time, separated by {action_sep}.
		Your reponse should be in the format of <think>...</think><answer>...</answer>
		e.g.
		<think>I can see from the sight the target object is right in the top left of me, I will move forward, then move left to access it.</think><answer>moveahead, moveahead,moveahead,moveahead,moveahead,moveleft,moveleft</answer>
		<think>I can see from the sight the target object is right in the top left of me, I will move forward, then move left to access it.</think><answer>moveahead{action_sep}moveahead{action_sep}moveahead{action_sep}moveahead{action_sep}moveahead{action_sep}moveleft{action_sep}moveleft</answer>
		"""

		system_prompt_vision = """You are a home robot and perform navigation tasks according to instructions.

vagen/env/primitive_skill/env.py

+28 −17

Original line number	Diff line number	Diff line
		@@ -26,40 +26,49 @@ class PrimitiveSkillEnv(BaseEnv):
		rst = parse_llm_raw_response(
		response=action_str,
		special_token_list=self.config.special_token_list,
		action_sep='\|',
		max_actions=1
		action_sep=self.config.action_sep,
		max_actions=self.config.max_actions_per_step,
		)
		output_info={}
		output_info.update(rst)
		action =self._parse_action(rst['actions'][0]) if len(rst['actions']) > 0 else None
		valid_action=rst['actions'][0] if len(rst['actions']) > 0 else ""
		valid_actions = []
		metrics = {
		"turn_metrics": {
		"action_is_valid": action is not None, # True if at least one valid action was parsed
		"action_is_valid": True, # True if at least one valid action was parsed
		},
		"traj_metrics": {
		"success": False, # Will be set to True if agent reaches goal
		},
		}
		if metrics["turn_metrics"]['action_is_valid']:
		reward += self.config.format_reward
		if action is not None:
		_, _, terminated, truncated, info = self.env.step(action)
		for action in rst['actions']:
		parsed_action = self._parse_action(action)
		if parsed_action is not None:
		_, _, terminated, truncated, info = self.env.step(parsed_action)
		valid_actions.append(action)
		self.last_info = info
		else:
		info=self.last_info
		terminated, truncated = False, False
		metrics["turn_metrics"]['action_is_valid'] = False
		break
		if truncated or terminated:
		break
		if metrics["turn_metrics"]['action_is_valid']:
		reward += self.config.format_reward
		if info['is_success']:
		metrics["traj_metrics"]['success'] = True
		done= terminated or truncated
		info["action_is_valid"] = action is not None
		obs=self._render(info,init_obs=False,valid_action=valid_action)
		info["action_is_valid"] = metrics["turn_metrics"]['action_is_valid']
		obs=self._render(info,init_obs=False,valid_actions=valid_actions)
		output_info["metrics"] = metrics
		self.total_reward += reward
		self.last_info = info
		return obs,reward,done,output_info

		def system_prompt(self):
		return system_prompt
		return system_prompt.format(
		max_actions_per_step=self.config.max_actions_per_step,
		action_sep=self.config.action_sep,
		)

		def close(self):
		self.env.close()
		@@ -84,7 +93,7 @@ class PrimitiveSkillEnv(BaseEnv):
		return (max_stage + 1) * 2+self.total_reward


		def _render(self,info,init_obs=False,valid_action=None):
		def _render(self,info,init_obs=False,valid_actions=None):
		new_info=handel_info(info.copy())
		object_positions=new_info['obj_positions']
		other_information=new_info['other_info']
		@@ -99,16 +108,18 @@ class PrimitiveSkillEnv(BaseEnv):
		other_information=other_information,
		x_workspace=x_workspace,
		y_workspace=y_workspace,
		z_workspace=z_workspace)
		z_workspace=z_workspace,
		max_action=self.config.max_actions_per_step)
		else:
		obs_str = action_template.format(valid_action=valid_action,
		obs_str = action_template.format(valid_actions=valid_actions,
		observation=img_placeholder,
		instruction=instruction,
		object_positions=object_positions,
		other_information=other_information,
		x_workspace=x_workspace,
		y_workspace=y_workspace,
		z_workspace=z_workspace)
		z_workspace=z_workspace,
		max_action=self.config.max_actions_per_step)
		multi_modal_data = None
		if self.config.render_mode == "vision":
		img=self.env.render()

Admin message