Gautam-R-Patil's picture
Watson upgrade: Four Pillars physics, score clamping (0.001-0.999), try/except safety
267dba3
Raw History Blame Contribute Delete
5.31 kB
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
"""
Data models for the Forensic Hawkeye Environment.
Forensic accident reconstruction environment where an AI agent acts as a
liability auditor, iteratively tuning physics parameters to reproduce crash
debris patterns and disprove false human testimony.
"""
import json
from typing import Dict, List, Optional, Tuple, Literal
from openenv.core.env_server.types import Action, Observation, State
from pydantic import Field, field_validator
class ForensicHawkeyeAction(Action):
"""
Action for the Forensic Hawkeye environment.
Two action types:
- RUN_SIMULATION: Provide physics parameters to simulate the crash
- SUBMIT_VERDICT: Submit a final liability determination
"""
action_type: Literal["RUN_SIMULATION", "SUBMIT_VERDICT"] = Field(
..., description="Type of action to perform"
)
thought: Optional[str] = Field(
default=None, description="Model's internal reasoning or thought process."
)
# Used for RUN_SIMULATION
sim_parameters: Optional[Dict[str, Dict[str, float]]] = Field(
default=None,
description='Physics parameters per entity, e.g. {"Car_A": {"speed": 45.0, "steering": 0.0}}',
)
# The 'Four Pillars' explicit parameters for RUN_SIMULATION
friction_coefficient: Optional[float] = Field(
default=None, description="Global friction coefficient (e.g. 0.3-0.9) deduced from weather"
)
restitution: Optional[float] = Field(
default=None, description="Global elasticity (e.g. 0.1-0.8) deduced from damage"
)
mass_overrides: Optional[Dict[str, float]] = Field(
default=None, description="Custom mass per entity in kg, e.g. {'Car_A': 8000.0}"
)
impact_offset_y: Optional[float] = Field(
default=None, description="Y-axis collision offset for angular momentum/spin"
)
# Used for SUBMIT_VERDICT
liable_party: Optional[str] = Field(
default=None, description="The entity determined to be at fault"
)
root_cause: Optional[str] = Field(
default=None, description="The root cause of the accident"
)
@field_validator("sim_parameters", mode="before")
@classmethod
def parse_sim_parameters(cls, v):
"""Auto-parse JSON strings from the web playground into dicts."""
if isinstance(v, str):
if not v.strip():
return None
return json.loads(v)
return v
class ForensicHawkeyeObservation(Observation):
"""
Observation from the Forensic Hawkeye environment.
Provides target debris positions (ground truth from the crash scene),
simulated debris positions (from the agent's last simulation), and
distance errors between them.
"""
# done: bool and reward: Optional[float] inherited from Observation
task_id: int = Field(default=1, description="Current task identifier (1-3)")
task_name: str = Field(default="", description="Human-readable task name")
target_debris: Dict[str, List[float]] = Field(
default_factory=dict,
description="Ground truth final positions from crash scene {entity: [x, y]}",
)
simulated_debris: Dict[str, List[float]] = Field(
default_factory=dict,
description="Simulated final positions from last run {entity: [x, y]}",
)
distance_errors: Dict[str, float] = Field(
default_factory=dict,
description="Per-entity distance between target and simulated positions",
)
total_distance_error: float = Field(
default=999.0,
description="Sum of all entity distance errors",
)
human_testimony: str = Field(
default="", description="The witness testimony to evaluate"
)
weather_conditions: str = Field(
default="", description="Extracted weather clues from testimony (for friction)"
)
vehicle_descriptions: Dict[str, str] = Field(
default_factory=dict, description="Extracted vehicle types (for mass)"
)
damage_description: str = Field(
default="", description="Extracted damage clues (for restitution)"
)
active_contradiction_flag: bool = Field(
default=False,
description="True if physics results contradict the human testimony",
)
step_number: int = Field(default=0, description="Current step in this episode")
max_steps: int = Field(default=15, description="Maximum steps allowed")
message: str = Field(default="", description="Feedback message from the environment")
available_entities: List[str] = Field(
default_factory=list,
description="Entities that can be configured in sim_parameters",
)
available_parameters: List[str] = Field(
default_factory=list,
description="Parameters that can be set per entity (e.g., speed, steering)",
)
class ForensicHawkeyeState(State):
"""
Internal state tracking for the Forensic Hawkeye environment.
"""
# episode_id and step_count inherited from State
task_id: int = 1
best_total_error: float = 999.0
verdict_submitted: bool = False
simulation_count: int = 0