Add natural language translation layer

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
+from gensim.utils import set_gpt_model, clear_messages, translate_natural_language_to_task_name
 
 
 class DemoRunner:
@@ -64,12 +64,21 @@ class DemoRunner:
         set_gpt_model(cfg['gpt_model'])
         cfg['load_memory'] = True
         cfg['use_template'] = True
-        cfg['target_task_name'] = target_task_name
+        
         cfg['task_description_candidate_num'] = 10
         cfg['record']['save_video'] = True
 
         print("cfg = ", cfg)
         memory = Memory(cfg)
+        
+        # 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...")
+            translated_name = translate_natural_language_to_task_name(target_task_name, interaction_txt=memory.chat_log)
+            print(f"Traduzido para: '{translated_name}'")
+            cfg['target_task_name'] = translated_name
+        else:
+            cfg['target_task_name'] = target_task_name
         agent = Agent(cfg, memory)
         critic = Critic(cfg, memory)
         self.simulation_runner = SimulationRunner(cfg, agent, critic, memory)
@@ -80,7 +89,15 @@ class DemoRunner:
 
     def run(self, instruction, progress):
         cfg = self._cfg
-        cfg['target_task_name'] = instruction
+        
+        # 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...")
+            translated_name = translate_natural_language_to_task_name(instruction, interaction_txt=self.simulation_runner.chat_log)
+            print(f"Traduzido para: '{translated_name}'")
+            cfg['target_task_name'] = translated_name
+        else:
+            cfg['target_task_name'] = instruction
 
         # self._env.cache_video = []
         self.simulation_runner._md_logger = ''
@@ -154,7 +171,7 @@ if __name__ == '__main__':
 
                 model_choice = gr.Radio(["gpt-3.5-turbo-16k", "gpt-4"], label="Qual modelo?", interactive=True)
                 option_choice = gr.Radio(["top-down"], label="Qual modo?", interactive=True)
-                inp_instruction = gr.Textbox(label='Nome da tarefa alvo (se usar top-down)', lines=1)
+                inp_instruction = gr.Textbox(label='Nome da tarefa alvo ou descrição em linguagem natural (ex: "construa uma ponte" ou "build-bridge")', lines=1)
                 info_setup = gr.Markdown(label='Status da configuração')
                 btn_setup = gr.Button("Configurar/Resetar simulação")
                 btn_run = gr.Button("Executar (pode levar 30+ segundos)")

gensim/utils.py

@@ -351,6 +351,46 @@ def clear_messages():
     existing_messages = []
 
 
+def translate_natural_language_to_task_name(natural_language_input, interaction_txt=None):
+    """
+    Converte linguagem natural (português ou inglês) para o formato de nome de tarefa esperado.
+    Exemplo: "construa uma ponte colorida" -> "build-colorful-bridge"
+    """
+    prompt = f"""Convert the following task description to a task name in the format "word-word-word" (lowercase, separated by hyphens).
+
+The task name should be concise and descriptive, following the pattern of existing tasks like:
+- "build-bridge" (for building a bridge)
+- "stack-blue-on-red" (for stacking blue blocks on red blocks)
+- "align-spheres-in-colored-zones" (for aligning spheres in colored zones)
+- "insert-sphere-into-container" (for inserting a sphere into a container)
+
+Input (in natural language): {natural_language_input}
+
+Return ONLY the task name in the format "word-word-word", nothing else. No explanations, no code blocks, just the task name."""
+    
+    try:
+        result = generate_feedback(prompt, max_tokens=50, temperature=0.0, interaction_txt=interaction_txt, retry_max=3)
+        # Limpar o resultado (remover espaços, quebras de linha, code blocks, etc)
+        task_name = result.strip()
+        # Remover markdown code blocks se houver
+        task_name = re.sub(r'```[a-z]*\n?', '', task_name).strip()
+        # Garantir que está em lowercase e com hífens
+        task_name = task_name.lower().replace(' ', '-').replace('_', '-')
+        # Remover caracteres inválidos
+        task_name = re.sub(r'[^a-z0-9-]', '', task_name)
+        
+        if interaction_txt is not None:
+            add_to_txt(interaction_txt, f"Translated '{natural_language_input}' to task name: {task_name}", with_print=True)
+        
+        return task_name
+    except Exception as e:
+        print(f"Error translating natural language: {e}")
+        # Fallback: tentar converter manualmente
+        fallback = natural_language_input.lower().replace(' ', '-').replace('_', '-')
+        fallback = re.sub(r'[^a-z0-9-]', '', fallback)
+        return fallback
+
+
 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)