[Fix] Update inference-only-demo for InternVLA-N1-DualVLN

internnav/agent/__init__.py

@@ -1,13 +1,8 @@
 from internnav.agent.base import Agent
 from internnav.agent.cma_agent import CmaAgent
+from internnav.agent.dialog_agent import DialogAgent
+from internnav.agent.internvla_n1_agent import InternVLAN1Agent
 from internnav.agent.rdp_agent import RdpAgent
 from internnav.agent.seq2seq_agent import Seq2SeqAgent
-from internnav.agent.internvla_n1_agent import InternVLAN1Agent
 
-__all__ = [
-    'Agent',
-    'CmaAgent',
-    'RdpAgent',
-    'Seq2SeqAgent',
-    'InternVLAN1Agent'
-]
+__all__ = ['Agent', 'DialogAgent', 'CmaAgent', 'RdpAgent', 'Seq2SeqAgent', 'InternVLAN1Agent']

internnav/agent/base.py

@@ -25,6 +25,7 @@ def decorator(agent_class):
             if agent_type in cls.agents:
                 raise ValueError(f"Agent {agent_type} already registered.")
             cls.agents[agent_type] = agent_class
+            return agent_class
 
         return decorator
 

internnav/agent/internvla_n1_agent.py

@@ -207,7 +207,7 @@ def s2_thread_func():
         self.s2_thread.daemon = True
         self.s2_thread.start()
 
-    def should_infer_s2(self, mode="sync"):
+    def should_infer_s2(self, mode="partial_async"):
         """Function: Enables the sys2 inference thread depending on the mode.
         mode: just support 2 modes: "sync" and "partial_async".
         "sync": Synchronous mode (navdp_version >= 0.0), Sys1 and Sys2 execute in a sequential inference chain.
@@ -298,8 +298,6 @@ def step(self, obs):
                 if self.sys1_infer_times > 0:
                     self.dual_forward_step += 1
 
-            # print('Output action:', output, self.dual_forward_step)
-
         else:
             self.look_down = False
             # 2. If output is in latent form, execute latent S1
@@ -333,13 +331,9 @@ def step(self, obs):
                         .unsqueeze(-1)
                         .to(self.device)
                     )  # [1, 2, 224, 224, 1]
-                    self.s1_output = self.policy.s1_step_latent(
-                        rgbs, depths, self.s2_output.output_latent, use_async=True
-                    )
+                    self.s1_output = self.policy.s1_step_latent(rgbs, depths, self.s2_output.output_latent)
                 else:
-                    self.s1_output = self.policy.s1_step_latent(
-                        rgb, depth * 10000.0, self.s2_output.output_latent, use_async=False
-                    )
+                    self.s1_output = self.policy.s1_step_latent(rgb, depth * 10000.0, self.s2_output.output_latent)
 
             else:
                 assert False, f"S2 output should be either action or latent, but got neither!  {self.s2_output}"
@@ -372,6 +366,7 @@ def step(self, obs):
                     if self.dual_forward_step > self.sys2_max_forward_step:
                         print("!!!!!!!!!!!!")
                         print("ERR: self.dual_forward_step ", self.dual_forward_step, " > ", self.sys2_max_forward_step)
+                        print("Potential reason: sys1 infers empty trajectory list []")
                         print("!!!!!!!!!!!!")
 
         print('Output discretized traj:', output['action'], self.dual_forward_step)

internnav/agent/internvla_n1_agent_realworld.py

@@ -222,10 +222,10 @@ def step_s2(self, rgb, depth, pose, instruction, intrinsic, look_down=False):
                 **inputs,
                 max_new_tokens=128,
                 do_sample=False,
-                use_cache=True,
-                past_key_values=self.past_key_values,
+                # use_cache=True,
+                # past_key_values=self.past_key_values,
                 return_dict_in_generate=True,
-                raw_input_ids=copy.deepcopy(inputs.input_ids),
+                # raw_input_ids=copy.deepcopy(inputs.input_ids),
             )
         output_ids = outputs.sequences
 
@@ -253,5 +253,5 @@ def step_s2(self, rgb, depth, pose, instruction, intrinsic, look_down=False):
             return action_seq, None, None
 
     def step_s1(self, latent, rgb, depth):
-        all_trajs = self.model.generate_traj(latent, rgb, depth, use_async=True)
+        all_trajs = self.model.generate_traj(latent, rgb, depth)
         return all_trajs

internnav/configs/evaluator/__init__.py

@@ -39,15 +39,17 @@ class MetricCfg(BaseModel):
 
 class TaskCfg(BaseModel):
     task_name: Optional[str] = None
-    task_settings: Dict[str, Any]
-    scene: SceneCfg
+    task_settings: Dict[str, Any] = None
+    scene: SceneCfg = None
     robot_name: Optional[str] = None
     robot: Optional[RobotCfg] = None
     robot_flash: Optional[bool] = None
+    flash_collision: Optional[bool] = None
     robot_usd_path: Optional[str] = None
     camera_resolution: Optional[List[int]] = None
     metric: Optional[MetricCfg] = None
     camera_prim_path: Optional[str] = None
+    one_step_stand_still: Optional[bool] = None
 
 
 class EvalDatasetCfg(BaseModel):

internnav/configs/evaluator/vln_default_config.py

@@ -239,6 +239,9 @@ def get_config(evaluator_cfg: EvalCfg):
 
     # add the flash controller in, by flash flag.
     if evaluator_cfg.task.robot_flash:
+        vln_move_by_flash_cfg.type = (
+            'VlnMoveByFlashCollisionController' if evaluator_cfg.task.flash_collision else 'VlnMoveByFlashController'
+        )
         robot.controllers.append(ControllerCfg(controller_settings=vln_move_by_flash_cfg.model_dump()))
 
     if evaluator_cfg.task.robot_flash or evaluator_cfg.eval_settings.get('vis_output', True):