FlyBrain-Lab / src /autonomy /engine.py
timfromhcs's picture
FlyBrain v4.1.0 Space build (REAL_SUBGRAPH, CPU-only, honest backend)
3d46076 verified
Raw
History Blame Contribute Delete
13.5 kB
"""Autonomy engine (STAGE E): self-generated goals, compositional actions,
self-evaluation. The organism decides what to inspect, where to go, what to
repeat and what to avoid — humans do not issue per-task commands.
Brain-driven contract (mission rule 13): candidates are MODULATED by neural
state (motor readout activations, drives, prediction error); the engine never
bypasses the brain — it composes continuous control from it.
"""
import math
from dataclasses import dataclass, field
from enum import Enum
from typing import Any, Dict, List, Optional, Tuple
import numpy as np
from src.common.determinism import derive_subseed
class GoalSource(str, Enum):
NEED = "need" # unmet homeostatic need (energy, safety)
CURIOSITY = "curiosity" # internal drive for novelty
PREDICTION_ERROR = "prediction_error"
OPPORTUNITY = "opportunity" # environmental (observed resource etc.)
MEMORY = "memory" # remembered success/failure
SOCIAL = "social"
class GoalStatus(str, Enum):
ACTIVE = "ACTIVE"
ACHIEVED = "ACHIEVED"
ABANDONED = "ABANDONED"
@dataclass
class Goal:
goal_id: str
source: str
kind: str # forage | explore | investigate | rest | socialize | avoid
target: Optional[Tuple[int, int]]
priority: float
created_tick: int
status: str = GoalStatus.ACTIVE.value
progress: float = 0.0
ticks_active: int = 0
evidence: Dict[str, Any] = field(default_factory=dict)
def to_dict(self) -> Dict[str, Any]:
d = dict(self.__dict__)
d["target"] = list(self.target) if self.target else None
return d
@dataclass
class ActionCandidate:
"""Compositional, continuous body control (NOT a tiny action menu)."""
kind: str # move | rest | investigate | communicate
heading: float # radians, continuous
speed: float # 0..1 continuous
duration: int # intended ticks
intensity: float # 0..1 e.g. communication/inspection strength
target: Optional[Tuple[int, int]] = None
goal_id: str = ""
neural_evidence: Dict[str, float] = field(default_factory=dict)
def to_dict(self) -> Dict[str, Any]:
d = dict(kind=self.kind, heading=round(self.heading, 4), speed=round(self.speed, 4),
duration=int(self.duration), intensity=round(self.intensity, 4),
target=list(self.target) if self.target else None,
goal_id=self.goal_id)
d["neural_evidence"] = {k: round(v, 5) for k, v in self.neural_evidence.items()}
return d
class AutonomyEngine:
"""Self-generated goal management + compositional action synthesis."""
def __init__(self, seed: int = 42, goal_patience: int = 12,
novelty_radius: int = 3):
self.seed = int(seed)
self.goal_patience = int(goal_patience)
self.novelty_radius = int(novelty_radius)
self.goals: List[Goal] = []
self.completed: List[Goal] = []
self._goal_seq = 0
self._visited: set = set()
self._rng_counter = 0
# ------------------------------------------------------------ goals
def _new_goal(self, tick: int, source: GoalSource, kind: str,
priority: float, target=None, evidence=None) -> Goal:
self._goal_seq += 1
gid = f"g{self.seed:04d}-{self._goal_seq:04d}"
g = Goal(gid, source.value, kind, target, round(float(priority), 4),
tick, evidence=dict(evidence or {}))
self.goals.append(g)
return g
def generate_goals(self, tick: int, energy: float, health: float,
drives: Dict[str, float], prediction_error: float,
world_sense: Dict[str, Any], position: Tuple[int, int],
skills: Dict[str, float]) -> List[Goal]:
"""Self-generated tasks from needs, drives, errors, opportunities, memory."""
generated: List[Goal] = []
# 1. Unmet needs dominate
if energy < 0.45:
generated.append(self._new_goal(
tick, GoalSource.NEED, "forage", priority=1.0 + (0.45 - energy),
evidence={"energy": round(energy, 3)}))
if health < 0.5:
generated.append(self._new_goal(
tick, GoalSource.NEED, "rest", priority=0.9 + (0.5 - health),
evidence={"health": round(health, 3)}))
# 2. Curiosity / novelty (unvisited neighborhood)
nearby = self._nearby_unvisited(position)
if drives.get("curiosity", 0.0) > 0.75 and nearby:
generated.append(self._new_goal(
tick, GoalSource.CURIOSITY, "explore", priority=0.3 + drives["curiosity"],
target=nearby[0], evidence={"unvisited_nearby": len(nearby)}))
# 3. Prediction error -> investigate
if prediction_error > 0.4:
generated.append(self._new_goal(
tick, GoalSource.PREDICTION_ERROR, "investigate",
priority=0.4 + prediction_error, target=position,
evidence={"prediction_error": round(prediction_error, 3)}))
# 4. Environmental opportunity
food = float(world_sense.get("food_gradient", 0.0))
if food > 1.0 and energy < 0.85:
generated.append(self._new_goal(
tick, GoalSource.OPPORTUNITY, "forage", priority=0.5 + min(1.0, food / 3.0),
evidence={"food_gradient": round(food, 3)}))
hazard = float(world_sense.get("hazard_gradient", 0.0))
if hazard > 0.6:
generated.append(self._new_goal(
tick, GoalSource.NEED, "avoid", priority=0.8 + hazard,
evidence={"hazard_gradient": round(hazard, 3)}))
# 5. Social drive
if drives.get("social", 0.0) > 0.9 and int(world_sense.get("nearby_organisms", 0)) > 0:
generated.append(self._new_goal(
tick, GoalSource.SOCIAL, "socialize", priority=0.35 + drives["social"],
evidence={"nearby": int(world_sense.get("nearby_organisms", 0))}))
return generated
def _nearby_unvisited(self, position: Tuple[int, int]) -> List[Tuple[int, int]]:
x, y = position
r = self.novelty_radius
cands = [(x + dx, y + dy) for dx in range(-r, r + 1) for dy in range(-r, r + 1)
if (dx, dy) != (0, 0)]
return [c for c in cands if c not in self._visited]
def update_goals(self, tick: int, energy_delta: float, novelty_gained: bool,
prediction_error: float) -> List[Goal]:
"""Self-evaluation: progress, achievement, abandonment, task switching."""
finished: List[Goal] = []
for g in self.goals:
g.ticks_active += 1
if g.kind == "forage":
g.progress = min(1.0, g.progress + max(0.0, energy_delta) * 2.0)
elif g.kind == "explore" and novelty_gained:
g.progress = min(1.0, g.progress + 0.34)
elif g.kind == "investigate":
g.progress = min(1.0, g.progress + max(0.0, 0.5 - prediction_error))
if g.progress >= 0.99:
g.status = GoalStatus.ACHIEVED.value
elif g.ticks_active > self.goal_patience and g.progress < 0.05:
g.status = GoalStatus.ABANDONED.value
if g.status != GoalStatus.ACTIVE.value:
finished.append(g)
if finished:
self.completed.extend(finished)
fin_ids = {g.goal_id for g in finished}
self.goals = [g for g in self.goals if g.goal_id not in fin_ids]
return finished
def active_goal(self) -> Optional[Goal]:
if not self.goals:
return None
return max(self.goals, key=lambda g: (g.priority, -g.created_tick))
def note_position(self, position: Tuple[int, int]) -> bool:
"""Track visited cells for novelty. Returns True if novel."""
if position in self._visited:
return False
self._visited.add(position)
return True
# ---------------------------------------------------------- actions
def synthesize_action(self, tick: int, goal: Optional[Goal],
brain_out: Dict[str, Any], world_sense: Dict[str, Any],
energy: float, health: float,
exploration: float = 0.5) -> ActionCandidate:
"""Compose continuous control from the ACTIVE goal + neural state.
Neural evidence (mission rule 13): motor_act asymmetry biases heading,
drive levels scale speed/intensity, prediction error boosts inspection.
"""
self._rng_counter += 1
rng = np.random.RandomState(derive_subseed(self.seed, f"act:{tick}:{self._rng_counter}"))
drives = brain_out.get("drives", {})
assoc = brain_out.get("tool_associations", {})
act_a = float(assoc.get("act_in_environment", 0.0))
remember_a = float(assoc.get("remember", 0.0))
pred_err = float(brain_out.get("prediction_error", 0.0))
# heading: goal target pull + neural lateral bias + deterministic jitter
base_heading = rng.uniform(0.0, 2.0 * math.pi)
if goal is not None and goal.target is not None:
dx, dy = goal.target[0] - world_sense["position"][0], goal.target[1] - world_sense["position"][1]
if abs(dx) + abs(dy) > 0:
base_heading = 0.4 * base_heading + 0.6 * math.atan2(dy, dx)
neural_bias = (act_a - remember_a) * (math.pi / 4.0)
heading = (base_heading + neural_bias) % (2.0 * math.pi)
kind = "move"
target = goal.target if goal is not None else None
if goal is not None:
if goal.kind == "rest":
kind = "rest"
elif goal.kind == "investigate":
kind = "investigate"
elif goal.kind == "socialize":
kind = "communicate"
elif health < 0.35:
kind = "rest"
curiosity = float(drives.get("curiosity", 0.5))
energy_drive = float(drives.get("energy", 1.0))
if kind == "rest":
speed = 0.0
duration = 3
intensity = float(np.clip(1.0 - health, 0.0, 1.0))
elif kind == "investigate":
speed = float(np.clip(0.2 + pred_err, 0.0, 1.0))
duration = 2
intensity = float(np.clip(pred_err, 0.0, 1.0))
elif kind == "communicate":
speed = float(np.clip(0.1 + 0.2 * drives.get("social", 0.5), 0.0, 1.0))
duration = 2
intensity = float(np.clip(drives.get("social", 0.5), 0.0, 1.0))
else:
# hungry organisms sprint; satiated organisms wander curiously
speed = float(np.clip(0.3 + (1.0 - energy) * 0.6 * energy_drive
+ 0.2 * curiosity * exploration, 0.05, 1.0))
duration = 1 + int(rng.randint(0, 2))
intensity = float(np.clip(act_a, 0.0, 1.0))
return ActionCandidate(
kind=kind, heading=float(heading), speed=float(np.clip(speed, 0.0, 1.0)),
duration=int(duration), intensity=float(np.clip(intensity, 0.0, 1.0)),
target=target, goal_id=goal.goal_id if goal is not None else "",
neural_evidence={
"motor_act": act_a, "motor_remember": remember_a,
"curiosity": curiosity, "prediction_error": pred_err,
})
# -------------------------------------------------------- introspection
def autonomy_summary(self) -> Dict[str, Any]:
return {
"active_goals": [g.to_dict() for g in self.goals],
"completed_count": len(self.completed),
"achieved": sum(1 for g in self.completed if g.status == GoalStatus.ACHIEVED.value),
"abandoned": sum(1 for g in self.completed if g.status == GoalStatus.ABANDONED.value),
"visited_cells": len(self._visited),
"goals_generated": self._goal_seq,
"sources": sorted({g.source for g in self.completed + self.goals}),
}
def snapshot(self) -> Dict[str, Any]:
return {
"seed": self.seed, "goal_patience": self.goal_patience,
"novelty_radius": self.novelty_radius,
"goals": [g.to_dict() for g in self.goals],
"completed": [g.to_dict() for g in self.completed],
"goal_seq": self._goal_seq,
"visited": sorted(list(self._visited)),
"rng_counter": self._rng_counter,
}
@classmethod
def restore(cls, payload: Dict[str, Any]) -> "AutonomyEngine":
eng = cls(payload["seed"], payload.get("goal_patience", 12),
payload.get("novelty_radius", 3))
def _goal(d):
t = d.get("target")
return Goal(d["goal_id"], d["source"], d["kind"],
tuple(t) if t else None, d["priority"], d["created_tick"],
d.get("status", "ACTIVE"), d.get("progress", 0.0),
d.get("ticks_active", 0), d.get("evidence", {}))
eng.goals = [_goal(d) for d in payload.get("goals", [])]
eng.completed = [_goal(d) for d in payload.get("completed", [])]
eng._goal_seq = int(payload.get("goal_seq", 0))
eng._visited = {tuple(c) for c in payload.get("visited", [])}
eng._rng_counter = int(payload.get("rng_counter", 0))
return eng