Remover camada de guardrail de validacao de input

app.py

@@ -22,7 +22,7 @@ from gensim.agent import Agent
 from gensim.critic import Critic
 from gensim.sim_runner import SimulationRunner
 from gensim.memory import Memory
-from gensim.utils import set_gpt_model, clear_messages, translate_natural_language_to_task_name, validate_task_input_with_llm
+from gensim.utils import set_gpt_model, clear_messages, translate_natural_language_to_task_name
 
 
 class DemoRunner:
@@ -71,20 +71,6 @@ class DemoRunner:
         print("cfg = ", cfg)
         memory = Memory(cfg)
         
-        # Validar input do usuário com guardrail LLM
-        if target_task_name and target_task_name.strip():
-            print(f"Validando input do usuário: '{target_task_name}'...")
-            is_valid, feedback, suggested_alternative = validate_task_input_with_llm(
-                target_task_name, 
-                interaction_txt=memory.chat_log
-            )
-            
-            if not is_valid:
-                error_msg = f"⚠️ Tarefa não é viável no simulador.\n\n{feedback}"
-                if suggested_alternative:
-                    error_msg += f"\n\n💡 Sugestão: {suggested_alternative}"
-                return error_msg, None
-        
         # Detectar se é linguagem natural e traduzir se necessário
         if target_task_name and (' ' in target_task_name or any(c.isupper() for c in target_task_name) or not target_task_name.replace('-', '').replace('_', '').isalnum()):
             print(f"Detectado linguagem natural: '{target_task_name}' - traduzindo...")
@@ -104,21 +90,6 @@ class DemoRunner:
     def run(self, instruction, progress):
         cfg = self._cfg
         
-        # Validar input do usuário com guardrail LLM
-        if instruction and instruction.strip():
-            print(f"Validando input do usuário: '{instruction}'...")
-            is_valid, feedback, suggested_alternative = validate_task_input_with_llm(
-                instruction, 
-                interaction_txt=self.simulation_runner.chat_log
-            )
-            
-            if not is_valid:
-                error_msg = f"⚠️ Tarefa não é viável no simulador.\n\n{feedback}"
-                if suggested_alternative:
-                    error_msg += f"\n\n💡 Sugestão: {suggested_alternative}"
-                yield f"Validação falhou: {error_msg}", error_msg, None, None
-                return
-        
         # Detectar se é linguagem natural e traduzir se necessário
         if instruction and (' ' in instruction or any(c.isupper() for c in instruction) or not instruction.replace('-', '').replace('_', '').isalnum()):
             print(f"Detectado linguagem natural: '{instruction}' - traduzindo...")
@@ -165,12 +136,7 @@ def setup(api_key, option_choice, model_choice, target_task_name):
     demo_runner = DemoRunner()
 
     if option_choice == 'top-down':
-        result = demo_runner.setup_top_down(api_key, target_task_name, model_choice)
-        # Se retornou tupla (erro, None), retornar diretamente
-        if isinstance(result, tuple) and len(result) == 2 and result[1] is None:
-            return result
-        # Caso contrário, é string de sucesso, concatenar com option_choice
-        info = result + option_choice
+        info = demo_runner.setup_top_down(api_key, target_task_name, model_choice) + option_choice
     # elif option_choice == 'bottom-up':
     #     info = demo_runner.setup(api_key, model_choice) + option_choice
     else:

gensim/utils.py

@@ -391,98 +391,6 @@ Return ONLY the task name in the format "word-word-word", nothing else. No expla
         return fallback
 
 
-def validate_task_input_with_llm(user_input, interaction_txt=None):
-    """
-    Valida o input do usuário usando LLM para verificar se a tarefa é viável
-    dentro das limitações do simulador e do framework.
-    
-    Returns:
-        (is_valid: bool, feedback: str, suggested_alternative: str)
-    """
-    prompt = f"""You are a validation system for a robotic simulation framework. Your job is to check if a user's task request is feasible within the system's limitations.
-
-SYSTEM LIMITATIONS:
-1. **Language Support**: The system accepts input in ANY language (Portuguese, English, Spanish, etc.). The input will be automatically translated to the task format. DO NOT reject tasks based on language.
-2. **Available Primitives**: Only PickPlace (pick and place) and Push are available. No cutting, breaking, throwing, or destructive actions.
-3. **Available Assets**: Only these assets exist:
-   - stacking/block.urdf (blocks)
-   - stacking/ball.urdf (balls/spheres)
-   - stacking/stand.urdf (stands/bases)
-   - container/container-template.urdf (containers)
-   - bowl/bowl.urdf (bowls)
-   - zone/zone.urdf (zones)
-   - corner/corner-template.urdf (corners)
-   - cylinder/cylinder-template.urdf (cylinders)
-   - box/box-template.urdf (boxes)
-   - pallet/pallet.urdf (pallets)
-   - No custom assets can be created
-4. **Workspace Bounds**: All objects must be within X: [0.25, 0.75], Y: [-0.5, 0.5], Z: [0, 0.3]
-5. **Object Limits**: Recommended 5-20 objects per task. More than 30-50 objects may cause performance issues.
-6. **Unsupported Actions**: 
-   - "Knock down" or "destroy" structures (can use "place near" or "push near" instead)
-   - Cutting, breaking, or modifying objects
-   - Throwing objects
-   - Complex physics interactions beyond pick-place-push
-   - Tasks requiring more than 25-30 steps
-7. **Available Colors**: red, blue, green, yellow, orange, purple, brown (no custom colors)
-8. **No Dynamic Object Creation**: All objects must be created at task initialization
-
-USER INPUT: "{user_input}"
-
-Analyze this task request and respond in the following JSON format:
-{{
-    "is_valid": true/false,
-    "reason": "brief explanation of why it's valid or invalid",
-    "violations": ["list of specific limitations violated, if any"],
-    "suggested_alternative": "suggested alternative task description if invalid, or empty string if valid"
-}}
-
-Respond ONLY with valid JSON, no other text."""
-
-    try:
-        result = generate_feedback(
-            prompt, 
-            max_tokens=300, 
-            temperature=0.0, 
-            interaction_txt=interaction_txt, 
-            retry_max=3
-        )
-        
-        # Limpar o resultado
-        result = result.strip()
-        result = re.sub(r'```json\s*', '', result)
-        result = re.sub(r'```\s*', '', result).strip()
-        
-        # Tentar parsear JSON
-        validation_result = json.loads(result)
-        
-        is_valid = validation_result.get('is_valid', False)
-        reason = validation_result.get('reason', 'No reason provided')
-        violations = validation_result.get('violations', [])
-        suggested_alternative = validation_result.get('suggested_alternative', '')
-        
-        # Construir feedback detalhado
-        feedback = reason
-        if violations:
-            feedback += "\n\nLimitações violadas:\n" + "\n".join(f"- {v}" for v in violations)
-        
-        if interaction_txt is not None:
-            status = "VALID" if is_valid else "INVALID"
-            add_to_txt(interaction_txt, f"Input validation ({status}): {reason}", with_print=True)
-        
-        return is_valid, feedback, suggested_alternative
-        
-    except json.JSONDecodeError as e:
-        print(f"Error parsing validation JSON: {e}")
-        print(f"Raw response: {result}")
-        # Fallback: assumir válido se não conseguir parsear
-        return True, "Validação não pôde ser processada, prosseguindo...", ""
-    except Exception as e:
-        print(f"Error in task validation: {e}")
-        # Fallback: assumir válido em caso de erro
-        return True, f"Erro na validação: {str(e)}. Prosseguindo...", ""
-
-
 def format_finetune_prompt(task_name):
     instruction_text = open('prompts/finetune_instructions_prompt.txt').read()
     instruction_text = instruction_text.replace("TASK_NAME_TEMPLATE", task_name)