---
title: Agents Perspective
emoji: 👀
colorFrom: red
colorTo: pink
sdk: gradio
sdk_version: 6.1.0
app_file: app.py
pinned: false
short_description: Experience the world through an RFT observer agent’s view
thumbnail: >-
  /static-proxy?url=https%3A%2F%2Fcdn-uploads.huggingface.co%2Fproduction%2Fuploads%2F685edcb04796127b024b4805%2FRfLfI-R1Q_Eq5_VzS9LO4.jpeg
---
# RFT Predator Space — Symmetric Observers (First-Person POV)

Experience the world from an RFT “observer agent” perspective inside a predator/prey arena rendered in pseudo-3D. The point is simple and testable: **both agents are symmetric observers co-existing in the same frame**, each with their own viewpoint, heading, and local policy.

This Space is intentionally inspectable: deterministic seeds, explicit rules, and no hidden model weights.

---

## Safety / Accessibility
**No flashing UI:** the “Map progression” panel is **event-driven** (updates only on reset/catch/load), not updated every timer tick. This prevents flicker that could be unsafe for photosensitive users.

---

## What this Space is
A deterministic gridworld where:
- **Predator** and **Prey** move on the same map.
- You view the world in **first-person** from the currently selected observer.
- The other agent is only visible when **line-of-sight + field-of-view** allow it (no “see through walls” cheating).
- Capture occurs if **Predator and Prey occupy the same cell**.

This is not a physics engine. It’s an observer-perception demo that’s easy to replay and verify.

---

## Core Concept: Symmetric Observers
Both agents have:
- position `(x, y)`
- orientation (E/S/W/N)
- local behavior rules

You can **toggle control** (Pred ↔ Prey) to swap whose “reality” you’re seeing. This is the cleanest way to demonstrate symmetry: neither observer is privileged.

---

## Modes (Manual / Auto / Hybrid)

### Manual (AutoRun OFF)
- Your button presses drive the currently controlled observer.
- The non-controlled agent still runs its own policy each step (so the world remains “alive”).

### AutoRun + AutoChase (AutoRun ON, AutoChase ON)
- Predator attempts to **chase** when the prey is within LOS+FOV.
- Otherwise predator **wanders**.
- Prey **flees** when not player-controlled.

### Hybrid AutoRun (AutoRun ON, AutoChase OFF)
- Predator **wanders autonomously** (no pursuit logic).
- Prey still **flees**.
- This demonstrates two independent observers operating in the same frame without collapsing into a single “predator narrative.”

**Important:** AutoRun moves the **current POV observer** (autopilot). If you’re viewing Predator POV, Predator moves under AutoRun. If you’re viewing Prey POV, Prey runs its flee autopilot.

---

## Rendering / “Seeing” Rules (First-Person POV)
The pseudo-3D view uses grid raycasting to draw walls and depth. The other agent only appears if:
1) **Line-of-sight** between the agents is clear  
2) The target is within the observer’s **field-of-view**  
3) The target is not hidden behind a closer wall slice (**occlusion**)

Result: the observer can’t see through walls, and perception is honest.

---

## Controls

### Movement (applies to the currently controlled observer)
- **Turn Left**
- **Forward**
- **Turn Right**

### Observer / Autonomy
- **Toggle Control (Pred ↔ Prey)**  
  Switches camera + manual input to the other observer.

- **Toggle AutoRun**  
  Enables autonomous ticking (timer based).

- **Toggle AutoChase**  
  Only affects predator behavior under AutoRun:
  - ON → chase policy (LOS + FOV) with roam fallback  
  - OFF → wander-only policy (Hybrid mode)

- **Tick**  
  Single deterministic step (useful for inspection and debugging).

### Symmetry tools
- **Swap Roles (Pred ⇄ Prey)**  
  Swaps positions and orientations. This is a “symmetry hammer” that makes it obvious both observers are equivalent in the frame.

### Optional overlay (kept subtle)
- **Toggle Overlay**  
  Displays faint “disturbance/coherence” visualization. It’s off by default to keep the base experience clean.

---

## Map Progression + Unlocks
Maps unlock as you accumulate catches. The map list shows:
- ✅ unlocked maps
- 🔒 locked maps and required catch count

Unlocks update only when something actually changes (reset/catch/load) to avoid flicker.

---

## Save / Load
Two supported workflows:

### Slot saves (server-side)
- Saves are written to `./saves/*.json`
- Dropdown lists existing saves (Refresh if needed)

### Export / Import (portable)
- Export produces a downloadable JSON save file
- Import lets you upload the save later to resume exactly

Saved state includes:
- seed, map name
- positions and orientations
- control target (Pred/Prey)
- mode toggles (AutoRun, AutoChase, overlay)
- catches (and therefore unlocked maps)

---

## Determinism / Reproducibility
Runs are deterministic given:
- the **Seed**
- the same sequence of actions / ticks

Autonomy uses deterministic RNG streams derived from seed + step index, so:
- same seed + same choices = same behavior

This is crucial for verification and repeatable demos.

---

## Files
- `app.py` — single-file Space for simplicity
- `saves/` — created automatically for slot saves

---

## Requirements
Minimal dependencies:
gradio
numpy

---

---

## Notes
This Space is designed to be inspectable and falsifiable at the “game rules” level: you can see exactly what the observer can see, why it turns, and when it commits to action.
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference