guychuk's picture
Upload src/models/agent.py
bc3d738 verified
"""
ARC-AGI-3 Agent: Full interactive agent combining JEPA, RSSM, and planning.
Core loop:
1. Encode observation via Grid-JEPA
2. Update RSSM world model with (obs, action) history
3. Use imagination rollouts to evaluate candidate actions
4. Execute best action in environment
5. Persist RSSM state across levels within an environment
6. TTT LoRA fine-tune on collected demos
7. Goal-inference from state transitions
"""
import random
from typing import List, Tuple, Optional, Dict
import torch
import torch.nn as nn
import torch.nn.functional as F
from encoder import GridPatchEmbed, ViTEncoder
from predictor import DiscreteActionEmbed, ActionConditionedPredictor
from grid_jepa import GridJEPA
from rssm import RSSM
class GoalInferenceModule(nn.Module):
"""Infers the goal/terminal state from observed transitions."""
def __init__(self, obs_dim: int, hidden_dim: int = 128):
super().__init__()
self.obs_dim = obs_dim
self.hidden_dim = hidden_dim
self.goal_encoder = nn.Sequential(
nn.Linear(obs_dim, hidden_dim), nn.ReLU(), nn.Linear(hidden_dim, hidden_dim),
)
self.goal_classifier = nn.Sequential(
nn.Linear(hidden_dim, 64), nn.ReLU(), nn.Linear(64, 1),
)
self.observed_goals: List[torch.Tensor] = []
def forward(self, obs_repr: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
goal_repr = self.goal_encoder(obs_repr)
is_goal_logit = self.goal_classifier(goal_repr)
return goal_repr, is_goal_logit
def register_terminal(self, obs_repr: torch.Tensor):
self.observed_goals.append(obs_repr.detach().cpu())
def get_goal_target(self) -> Optional[torch.Tensor]:
if len(self.observed_goals) == 0:
return None
return torch.stack(self.observed_goals).mean(dim=0)
class UncertaintyTracker:
"""Tracks prediction errors to detect when the world model is wrong and triggers hypothesis revision."""
def __init__(
self,
window_size: int = 5,
error_threshold: float = 2.0,
revision_threshold: int = 3,
):
self.window_size = window_size
self.error_threshold = error_threshold
self.revision_threshold = revision_threshold
self.prediction_errors: List[float] = []
self.revision_count: int = 0
self.last_revision_step: int = 0
def record_prediction_error(self, predicted_obs: torch.Tensor, actual_obs: torch.Tensor):
error = torch.norm(predicted_obs - actual_obs).item()
self.prediction_errors.append(error)
if len(self.prediction_errors) > self.window_size:
self.prediction_errors.pop(0)
def should_revise_hypothesis(self) -> bool:
if len(self.prediction_errors) < self.revision_threshold:
return False
recent_errors = self.prediction_errors[-self.revision_threshold:]
high_error_count = sum(1 for e in recent_errors if e > self.error_threshold)
return high_error_count >= self.revision_threshold
def get_error_stats(self) -> Dict[str, float]:
if len(self.prediction_errors) == 0:
return {"mean": 0.0, "max": 0.0, "recent": 0.0, "revision_count": self.revision_count}
recent = self.prediction_errors[-self.window_size:]
return {
"mean": sum(self.prediction_errors) / len(self.prediction_errors),
"max": max(self.prediction_errors),
"recent": sum(recent) / len(recent),
"revision_count": self.revision_count,
}
def mark_revision(self, step: int):
self.revision_count += 1
self.last_revision_step = step
self.prediction_errors.clear()
class ExplorationPolicy:
"""Novelty-seeking exploration for unknown ARC environments."""
def __init__(self, num_actions: int, grid_size: int = 64):
self.num_actions = num_actions
self.grid_size = grid_size
self.num_positions = grid_size * grid_size
self.visited_states: set = set()
self.action_history: List[Tuple[int, int]] = []
def hash_state(self, grid: torch.Tensor) -> int:
return hash(grid.cpu().numpy().tobytes())
def select_action(self, grid: torch.Tensor, novelty_bonus: bool = True, avoid_undo: bool = True) -> Tuple[int, int]:
state_hash = self.hash_state(grid)
is_novel = state_hash not in self.visited_states
self.visited_states.add(state_hash)
action_key = random.randint(0, self.num_actions - 1)
grid_np = grid.cpu().numpy()
import numpy as np
non_bg = list(zip(*np.where(grid_np != 0)))
if len(non_bg) > 0 and random.random() < 0.7:
r, c = random.choice(non_bg)
action_pos = r * self.grid_size + c
else:
action_pos = random.randint(0, self.num_positions - 1)
if avoid_undo and len(self.action_history) > 0:
last_key, last_pos = self.action_history[-1]
if action_key == last_key and action_pos == last_pos:
action_key = (action_key + 1) % self.num_actions
self.action_history.append((action_key, action_pos))
return action_key, action_pos
def reset(self):
self.visited_states.clear()
self.action_history.clear()
class PlanningModule:
"""Model-based planning using RSSM imagination rollouts."""
def __init__(
self,
rssm: RSSM,
goal_module: GoalInferenceModule,
jepa_encoder: GridJEPA,
horizon: int = 10,
num_candidates: int = 16,
):
self.rssm = rssm
self.goal_module = goal_module
self.jepa_encoder = jepa_encoder
self.horizon = horizon
self.num_candidates = num_candidates
def plan_action(
self, grid: torch.Tensor, h_state: torch.Tensor, z_state: torch.Tensor,
num_actions: int, device: torch.device,
) -> Tuple[int, int]:
B = 1
obs_repr = self.jepa_encoder.encode(grid)
obs_repr = obs_repr.mean(dim=1)
goal_target = self.goal_module.get_goal_target()
num_positions = grid.shape[-1] * grid.shape[-2]
total_actions = num_actions * num_positions
candidate_actions = torch.randint(0, total_actions, (B, self.num_candidates, self.horizon), device=device)
best_score = float("-inf")
best_action_idx = 0
for i in range(self.num_candidates):
actions = candidate_actions[0, i]
h_roll, z_roll = h_state.clone(), z_state.clone()
rollout_scores = []
for t in range(self.horizon):
a = actions[t:t+1]
h_roll, z_roll, _ = self.rssm.imagine(h_roll, z_roll, a)
if goal_target is not None:
dist_to_goal = -torch.norm(z_roll - goal_target.to(device))
rollout_scores.append(dist_to_goal.item())
else:
continue_logits = self.rssm.predict_continue(h_roll, z_roll)
rollout_scores.append(-torch.sigmoid(continue_logits).item())
avg_score = sum(rollout_scores) / len(rollout_scores) if rollout_scores else 0.0
if goal_target is not None and avg_score > -0.1:
avg_score += (self.horizon - len(rollout_scores)) * 0.1
if avg_score > best_score:
best_score = avg_score
best_action_idx = i
best_action = candidate_actions[0, best_action_idx, 0].item()
action_key = best_action // num_positions
action_pos = best_action % num_positions
return action_key, action_pos
class ARCAgent(nn.Module):
"""Complete ARC-AGI-3 agent with persistent state across levels."""
def __init__(
self,
jepa: GridJEPA,
rssm: RSSM,
num_actions: int = 6,
grid_size: int = 64,
exploration_ratio: float = 0.3,
device: str = "cuda",
):
super().__init__()
self.jepa = jepa
self.rssm = rssm
self.num_actions = num_actions
self.grid_size = grid_size
self.num_positions = grid_size * grid_size
self.exploration_ratio = exploration_ratio
self.device = torch.device(device if torch.cuda.is_available() else "cpu")
obs_dim = jepa.embed_dim
self.goal_module = GoalInferenceModule(obs_dim)
self.exploration = ExplorationPolicy(num_actions, grid_size)
self.planning = PlanningModule(rssm, self.goal_module, jepa)
self.uncertainty_tracker = UncertaintyTracker()
self.persistent_h: Optional[torch.Tensor] = None
self.persistent_z: Optional[torch.Tensor] = None
self.demo_buffer: List[Dict] = []
self.step_counter: int = 0
def reset_for_new_environment(self):
"""Reset ALL state when starting a completely new environment/game."""
self.persistent_h = None
self.persistent_z = None
self.exploration.reset()
self.goal_module.observed_goals.clear()
self.demo_buffer.clear()
self.step_counter = 0
self.uncertainty_tracker = UncertaintyTracker()
def reset_for_new_level(self):
"""Reset level-specific state but PERSIST world model knowledge."""
self.exploration.reset()
# DO NOT reset persistent_h/persistent_z
def encode_observation(self, grid: torch.Tensor) -> torch.Tensor:
return self.jepa.encode(grid)
def step(
self,
grid: torch.Tensor,
reward: Optional[float] = None,
done: bool = False,
is_exploration_phase: bool = False,
) -> Tuple[int, int]:
grid = grid.to(self.device)
obs_repr = self.encode_observation(grid)
obs_repr_pooled = obs_repr.mean(dim=1)
if self.persistent_h is None:
self.persistent_h, self.persistent_z = self.rssm.init_state(1, self.device)
if len(self.exploration.action_history) == 0:
prev_action = torch.zeros(1, dtype=torch.long, device=self.device)
else:
last_key, last_pos = self.exploration.action_history[-1]
prev_action = torch.tensor([last_key * self.num_positions + last_pos], device=self.device)
self.persistent_h, self.persistent_z, _, _ = self.rssm.observe(
obs_repr_pooled, prev_action, self.persistent_h, self.persistent_z
)
if done:
self.goal_module.register_terminal(obs_repr_pooled)
# Check if we need hypothesis revision
if self.uncertainty_tracker.should_revise_hypothesis():
# Reset exploration to try new strategies
self.exploration.reset()
self.uncertainty_tracker.mark_revision(self.step_counter)
if is_exploration_phase or random.random() < self.exploration_ratio:
action_key, action_pos = self.exploration.select_action(grid[0])
else:
action_key, action_pos = self.planning.plan_action(
grid, self.persistent_h, self.persistent_z, self.num_actions, self.device
)
self.demo_buffer.append({
"grid": grid[0].cpu().clone(),
"action_key": action_key,
"action_pos": action_pos,
"obs_repr": obs_repr_pooled.detach().cpu().clone(),
"h_state": self.persistent_h.detach().cpu().clone(),
"z_state": self.persistent_z.detach().cpu().clone(),
})
self.step_counter += 1
return action_key, action_pos
def run_level(self, env, max_steps: int = 100, exploration_steps: int = 10) -> Dict:
trajectory = []
for step_idx in range(max_steps):
grid = env.get_observation().unsqueeze(0)
reward, done = env.get_reward(), env.is_done()
is_exploration = step_idx < exploration_steps
action_key, action_pos = self.step(grid, reward, done, is_exploration)
env.step(action_key, action_pos)
trajectory.append({
"step": step_idx,
"action_key": action_key,
"action_pos": action_pos,
"reward": reward,
"done": done,
})
if done:
break
return {"trajectory": trajectory, "num_steps": len(trajectory), "success": done}
def create_agent(num_colors: int = 16, embed_dim: int = 384, grid_size: int = 64,
num_actions: int = 6, device: str = "cuda") -> ARCAgent:
jepa = GridJEPA(num_colors=num_colors, embed_dim=embed_dim, encoder_depth=12,
predictor_depth=12, num_heads=6, max_grid_size=grid_size)
rssm = RSSM(embed_dim=embed_dim, latent_dim=32, latent_classes=32, hidden_dim=256,
action_dim=64, num_actions=num_actions * grid_size * grid_size, obs_dim=embed_dim)
agent = ARCAgent(jepa=jepa, rssm=rssm, num_actions=num_actions, grid_size=grid_size, device=device)
agent = agent.to(device)
return agent
if __name__ == "__main__":
import numpy as np
device = "cuda" if torch.cuda.is_available() else "cpu"
agent = create_agent(num_colors=10, embed_dim=192, grid_size=10, num_actions=6, device=device)
class MockEnv:
def __init__(self, size=10):
self.grid = torch.zeros(size, size, dtype=torch.long)
self.grid[size//2, size//2] = 1
self.step_count = 0
self.max_steps = 20
def get_observation(self):
return self.grid
def get_reward(self):
return 0.0
def is_done(self):
return self.step_count >= self.max_steps
def step(self, action_key, action_pos):
r = action_pos // self.grid.shape[0]
c = action_pos % self.grid.shape[0]
if 0 <= r < self.grid.shape[0] and 0 <= c < self.grid.shape[1]:
self.grid[r, c] = action_key
self.step_count += 1
env = MockEnv(size=10)
grid = env.get_observation().unsqueeze(0).to(device)
action_key, action_pos = agent.step(grid)
print(f"Action: key={action_key}, pos={action_pos}")
agent.reset_for_new_environment()
result = agent.run_level(env, max_steps=15, exploration_steps=5)
print(f"Level result: {result['num_steps']} steps, success={result['success']}")
h_before = agent.persistent_h.clone() if agent.persistent_h is not None else None
env2 = MockEnv(size=10)
agent.reset_for_new_level()
result2 = agent.run_level(env2, max_steps=10, exploration_steps=3)
h_after = agent.persistent_h.clone() if agent.persistent_h is not None else None
if h_before is not None and h_after is not None:
state_persisted = not torch.allclose(h_before, torch.zeros_like(h_before))
print(f"State persisted across levels: {state_persisted}")
print("\nAgent tests passed!")