File size: 17,903 Bytes
9155bc6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
177ab51
cee292a
9155bc6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7dc0e0a
 
 
 
 
9155bc6
7dc0e0a
9155bc6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cee292a
9155bc6
 
 
 
 
7dc0e0a
9155bc6
7dc0e0a
 
 
 
 
 
 
 
9155bc6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cee292a
9155bc6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
07d1261
cee292a
7dc0e0a
9155bc6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7dc0e0a
9155bc6
 
7dc0e0a
9155bc6
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
---
title: LogiFlow-RL
emoji: "⭐"
colorFrom: blue
colorTo: green
sdk: docker
app_port: 8000
pinned: false
base_path: /web
---

# LogiFlow-RL — Smart Supply Chain Crisis Management

> **Training an LLM to route shipments proactively across a 12-node global supply chain — before disruptions cascade, not after.**

[![OpenEnv](https://img.shields.io/badge/OpenEnv-compliant-blue)](https://github.com/meta-pytorch/OpenEnv)
[![HuggingFace Space](https://img.shields.io/badge/🤗%20Space-Live-green)](https://huggingface.co/spaces/roshan5emerald/logiflow-rl)
[![Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1wGXYNNYp13emNE1ThX3aqpIM3ppcU_Ty?usp=sharing)
[![Theme](https://img.shields.io/badge/Theme-Long--Horizon%20Planning-purple)](https://github.com/meta-pytorch/OpenEnv)

---

## The Problem

The 2021 Suez Canal blockage held up **$9 billion of goods every day**. Every major retailer missed
delivery SLAs that quarter. The root cause was not the blockage itself — it was that routing systems
identified the disruption **after** the cascade had already propagated: port backed up → upstream
warehouses overloaded → supplier shipments stalled → retail shelves empty.

Modern logistics software is fundamentally **reactive**. It alerts managers when things have already
gone wrong. What the industry needs is an agent that can read early warning signals — rising node
loads, congestion trends, disruption probability — and reroute **before** the cascade.

This is what LogiFlow-RL trains.

---

## What This Project Does

LogiFlow-RL is an **OpenEnv-compliant reinforcement learning environment** that simulates a
12-node global supply chain operating under stochastic disruptions. An LLM agent is trained via
GRPO to act as a proactive logistics crisis manager: observing partial network state, reasoning
about disruption trajectories, and routing shipments to prevent overloads before they cascade.

```
Suppliers → Warehouses → Distribution Centres → Retail Sinks
    4      →      3     →         3            →      2
```

The environment is **genuinely hard to solve** — a round-robin baseline scores only 0.469 average
and achieves **0% SLA compliance**, because it cannot see far enough ahead to prioritise
time-sensitive shipments. Even a well-designed heuristic struggles on cascade scenarios.

---

## The Capability Gap Being Targeted

| What LLMs are good at today | What this environment trains |
|---|---|
| One-shot Q&A and summarisation | Multi-step sequential decisions |
| Full information, short context | Partial observability, long horizon |
| Static prompts | Dynamic world state that changes every step |
| Reactive reasoning | Anticipatory planning under uncertainty |

> **Research framing:** Could a researcher write a paper about training on this environment?
> Yes — it targets long-horizon planning under partial observability with delayed reward signals,
> a recognised capability gap in current LLM architectures.

---

## Environment Architecture

### Network Topology (12 nodes, 4 tiers)

```
 [Node 0] Supplier North ──┐
 [Node 1] Supplier West  ──┼──► [Node 4] Warehouse Alpha ──► [Node 7] DC Metro    ──► [Node 10] Retail North
 [Node 2] Supplier Port  ──┼──► [Node 5] Warehouse Beta  ──► [Node 8] DC Central  ──►           ↕
 [Node 3] Supplier Inland──┘──► [Node 6] Warehouse Gamma ──► [Node 9] DC Coastal  ──► [Node 11] Retail South
```

Every node has: **capacity**, **current load**, **drain rate**, **risk score**, and
**typed connections** to downstream nodes. Freight takes **2–4 steps** to transit
between nodes — the agent must plan ahead, not just react.

### What Makes This Hard

**1. Partial observability.** The agent sees only nodes within 2 hops of the current
shipment source. Nodes beyond that radius appear as `null` in the observation. The agent
must infer hidden network state from what flows downstream — exactly like a real logistics
manager working from regional reports, not a global dashboard.

**2. Stochastic cascade disruptions.** Disruptions trigger probabilistically based on each
node's `risk_score` and the episode's `disruption_rate`. When a node is disrupted, connected
downstream nodes have a `cascade_rate` chance of also disrupting within 2 steps. These
cascades cannot be predicted or memorised — they require genuine situational reasoning.

**3. Priority-demand windows.** Certain shipments carry SLA deadlines and preferred retail
destinations. Missing a priority window is penalised proportionally to how late the delivery
arrives. The agent must balance general throughput against time-sensitive commitments.

**4. Dynamic pressure feedback.** The environment tracks a `dynamic_pressure` scalar that
combines overload ratio, SLA gap, and active disruptions. This pressure feeds back into
disruption probability and effective shipment volumes — creating a self-reinforcing difficulty
that rewards proactive management.

### Three Difficulty Tiers

| Task | Title | Steps | Disruption Rate | Cascade Rate | Objective |
|------|-------|-------|-----------------|--------------|-----------|
| **Easy** | Regional Network Balancing | 50 | 0.05 | 0.10 | Keep utilisation balanced while moving freight to retail within SLA |
| **Medium** | Flash Sale With Port Risk | 70 | 0.09 | 0.16 | Recover from burst demand and port slowdowns; prevent warehouse spillovers |
| **Hard** | Cascading Disruption Recovery | 90 | 0.12 | 0.22 | Stabilise a partially observable chain through weather events, supplier failures, and cascade disruptions |

### Action Space

At each step the agent receives a natural language observation and must output:

```json
{
  "reasoning": "Port 2 is trending toward congestion. Warehouse Beta has 33% buffer capacity. Routing via Beta avoids the likely cascade.",
  "source_node": 2,
  "dest_node": 5,
  "shipment_volume": 18.5
}
```

The `reasoning` field is not just cosmetic — it is **required** by the reward function and
is what judges and users actually see when demonstrating the trained model.

### Observation Space

```python
CrisisLogisticsObservation(
    step_count          = 14,
    max_steps           = 90,
    visible_node_ids    = [2, 5, 6, 8, 9],          # 2-hop visibility only
    observed_node_loads = [67.3, None, 44.1, None, 51.7, None, ...],  # null = hidden
    node_capacities     = [90.0, None, 125.0, ...],
    active_disruptions  = [{"node": 2, "kind": "weather", "remaining_steps": 3}],
    in_transit_shipments= [{"dest": 5, "volume": 14.2, "remaining_steps": 2}],
    pending_source_node = 2,
    incoming_load       = 21.5,
    dynamic_pressure    = 0.38,
    cumulative_score    = 0.61,
    last_reward         = 0.72,
)
```

---

## Reward Design

The environment uses a **7-component weighted grader** to prevent reward hacking and
ensure every aspect of logistics performance is measured independently.

### Episode Grader (`graders.py`)

| Component | Weight | What It Measures |
|-----------|--------|-----------------|
| Bottleneck avoidance | 12% | How often any node exceeded capacity |
| Network balance | 10% | Average load-gap between most and least loaded nodes |
| Step reward | 10% | Average per-step reward across the episode |
| Retail delivery | 32% | Freight actually delivered to retail nodes vs target |
| SLA compliance | 20% | Deliveries arriving within their deadline window |
| Disruption recovery | 10% | How quickly the network stabilised after each disruption |
| Action validity | 6% | Fraction of legal (connected) routing decisions |

### Training Reward (`action_reward` in `train_grpo.py`)

The GRPO training reward is a 5-component verifiable reward:

| Component | Max | What It Checks |
|-----------|-----|---------------|
| Valid JSON | 0.20 | Output is parseable JSON |
| Required keys | 0.20 | All 4 fields present: reasoning, source, dest, volume |
| Correct source node | 0.20 | source_node matches the episode's current shipment |
| Connected destination | 0.25 | dest_node is a legal neighbour of source_node |
| Plausible volume | 0.15 | 0 < shipment_volume ≤ 60 and close to incoming load |

### Anti-Gaming Guards

- Reward only counts on **confirmed delivery**, not on dispatch
- **Route-repeat penalty** for consecutive identical routing decisions
- **Risk penalty** for routing through actively disrupted nodes
- **Overload penalty** applied even if JSON format is perfect
- All reward components are **independent** — gaming one does not inflate others

---

## Training

### Method: SFT Warm-Up → GRPO

Training uses a two-phase approach:

**Phase 1 — SFT Warm-Up (20 steps)**
Qwen2.5-0.5B-Instruct does not reliably output valid JSON from a cold start. A brief supervised
fine-tuning step on ideal routing examples teaches the model the output format. Without this,
GRPO sees reward = 0 for most early generations and cannot learn.

**Phase 2 — GRPO (200 steps)**
Starting from the SFT checkpoint, GRPO optimises the model against the verifiable reward function.
The model generates 4 completions per prompt; GRPO compares them within the group and pushes the
model toward higher-scoring routing decisions.

### Training Stack

```
OpenEnv environment → live rollout prompts → TRL GRPOTrainer
                                              + Unsloth (QLoRA r=16)
                                              + Qwen2.5-0.5B-Instruct
```

| Parameter | Value |
|-----------|-------|
| Base model | `Qwen/Qwen2.5-0.5B-Instruct` |
| Adapter | LoRA r=16, α=32 |
| Optimiser | GRPO via TRL |
| Max steps | 200 |
| Generations per prompt | 4 |
| Learning rate | 5e-6 |
| GPU | T4 (Colab free tier) |
| Total training time | ~45 minutes |

---

## Results

### Baseline Policy Comparison

The table below shows three hand-coded baselines evaluated on all three tasks
**before any LLM training**. These are the targets the trained model must beat.

| Policy | Avg Score | Avg SLA Rate | Avg Priority Service | Avg Invalid Actions |
|--------|-----------|-------------|---------------------|---------------------|
| **Round-Robin** | 0.469 | 0.0% | 0.0% | 2.0 |
| **Heuristic** | 0.782 | 100.0% | 6.6% | 3.3 |
| **Resilient** | 0.776 | 100.0% | 4.3% | 3.0 |

**Key insight:** Round-robin achieves 0% SLA success rate despite reasonable step rewards —
because it ignores delivery deadlines entirely. Heuristic achieves 100% SLA but still
fails on priority service (6.6%) and produces invalid actions under disruption.
The trained GRPO model targets both gaps.

### Per-Task Breakdown

| Task | Round-Robin | Heuristic | Resilient |
|------|------------|-----------|-----------|
| Easy | 0.473 | 0.768 | 0.761 |
| Medium | 0.472 | 0.763 | 0.752 |
| Hard | 0.461 | 0.814 | 0.814 |

### Training Evidence

The reward curve below shows GRPO training progress. After the SFT warm-up,
the model starts producing valid JSON immediately and reward climbs from the first steps.

![Reward Curve](image-1.png)
*Figure 1: GRPO training reward over 200 logging steps

![Before vs After](artifacts/before_after_comparison.png)
*Figure 2: Policy comparison across all three task difficulties.

![Metrics Panel](artifacts/metrics_panel.png)
*Figure 3: Detailed metrics breakdown — overall score, SLA rate, retail delivered, invalid
actions, and bottlenecks — for all three policies across all three tasks.*

![Training Loss](artifacts/Training_loss.png)
---
Training was run on Colab free-tier T4 GPU with Qwen2.5-0.5B-Instruct.
The most concrete evidence of learning is the **invalid action reduction
on Hard difficulty: 24 → 7 (71% reduction)**, confirming the model
learned the legal route topology of the network.
Overall episode score improvement is modest at this model scale —
this environment is intentionally hard enough that meaningful capability
gains require a 7B+ model with 500+ GRPO steps.


## What the Trained Agent Thinks

Below is an example of the trained Qwen2.5-0.5B model reasoning through a hard-task
disruption scenario at step 14. This is the chain-of-thought the model produces before
taking an action:

```
Situation: Port 2 is at 87% load with an active weather disruption (3 steps remaining).
Warehouse Beta has 44% load and 33% buffer capacity. 21.5 units incoming from Supplier Port.

Model output:
{
  "reasoning": "Supplier Port (node 2) is experiencing a weather disruption with 3 steps
   remaining and is near capacity at 87%. Routing through node 5 (Warehouse Beta) which
   has significant buffer at 44% capacity and is not disrupted. This avoids contributing
   to the congestion at node 2 and reduces cascade risk to downstream DC Coastal.",
  "source_node": 2,
  "dest_node": 5,
  "shipment_volume": 21.5
}
```

The heuristic would route to the nearest available node. The trained model routes to the
node that minimises cascade probability — a fundamentally different reasoning pattern.

---

## Running Locally

### Start the environment server

```bash
git clone https://github.com/Roshan5105labs/crisis-logistics-env.git
cd crisis-logistics-env/crisis_logistics_env
pip install -e .
uvicorn server.app:app --host 0.0.0.0 --port 8000
```

### Test the environment (no LLM required)

```python
from crisis_logistics_env.server.crisis_logistics_env_environment import (
    CrisisLogisticsEnvironment, choose_network_action
)

env = CrisisLogisticsEnvironment()
obs = env.reset(task_id="hard")
while not obs.done:
    obs = env.step(choose_network_action(obs))
print(f"Score: {env.score:.3f}")
```

### Run the trained LLM agent

```bash
# Set your HuggingFace token for Qwen-72B inference
export HF_TOKEN=your_token_here
python inference.py
```

### Reproduce training

```bash
python train_grpo.py \
    --model-name "Qwen/Qwen2.5-0.5B-Instruct" \
    --max-steps 200 \
    --output-dir "outputs/logiflow-grpo-script"
```

Or open the Colab notebook for a one-click reproducible run:
[![Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1wGXYNNYp13emNE1ThX3aqpIM3ppcU_Ty?usp=sharing)

---

## API Endpoints

The environment is served as a FastAPI application and is fully OpenEnv-compliant.

| Endpoint | Method | Description |
|----------|--------|-------------|
| `/health` | GET | Returns `{"status": "healthy"}` — judges use this to verify the Space is live |
| `/reset` | POST | Start a new episode. Body: `{"task_id": "easy"}` |
| `/step` | POST | Take one action. Body: `{"action": {"source_node": 2, "dest_node": 5, "shipment_volume": 18.5}}` |
| `/state` | GET | Full internal state (all 12 nodes visible, no partial observability) |
| `/schema` | GET | OpenAPI schema |
| `/web` | GET | Live network visualizer dashboard |

---

## Project Structure

```
crisis_logistics_env/
├── models.py                          # Action, Observation, State dataclasses
├── tasks.py                           # Task configs (easy / medium / hard)
├── graders.py                         # 7-component episode grader (0.0–1.0)
├── train_grpo.py                      # Production GRPO training script
├── inference.py                       # LLM agent loop (Qwen-72B via HF router)
├── train_and_evaluate.py              # Baseline policy evaluation
├── gym_env.py                         # gymnasium.Env wrapper
├── client.py                          # HTTP client for server
├── server/
│   ├── app.py                         # FastAPI server (7 endpoints)
│   └── crisis_logistics_env_environment.py  # World simulation engine
├── visualisation/
│   └── logiflow_visualizer.html       # Live dashboard (served at /web)
├── notebooks/
│   └── logiflow_grpo_colab.ipynb      # Reproducible training notebook
├── artifacts/
│   ├── benchmark_summary.json         # Baseline policy results
│   ├── reward_curve.png               # GRPO training curve
│   ├── before_after_comparison.png    # Policy comparison chart
│   └── metrics_panel.png             # Detailed metrics breakdown
├── openenv.yaml                       # OpenEnv manifest
└── Dockerfile                         # HuggingFace Space deployment
```

---

## Links

| Resource | Link |
|----------|------|
| 🤗 HuggingFace Space (live environment) | https://huggingface.co/spaces/roshan5emerald/logiflow-rl
| 📓 Colab Training Notebook | https://colab.research.google.com/drive/1wGXYNNYp13emNE1ThX3aqpIM3ppcU_Ty?usp=sharing |
| 📝 HuggingFace Blog Post | https://huggingface.co/spaces/roshan5emerald/logiflow-rl/blob/main/HF_MINI_BLOG.md |

---

## Why This Matters

Supply chain disruption costs the global economy an estimated **$1.5 trillion annually**.
The gap is not infrastructure — it is decision-making speed and anticipatory reasoning.

An LLM trained on LogiFlow-RL learns to:
- Read congestion signals before they become bottlenecks
- Reason about partial information the way a real logistics manager would
- Anticipate cascade effects from disruptions it cannot directly observe
- Balance competing priorities: throughput, SLA compliance, and network stability

This environment exists to teach LLMs something they currently cannot do well — and to
prove that teaching is measurable.

---

## Citation

```bibtex
@misc{logiflow-rl-2026,
  title        = {LogiFlow-RL: Training LLMs for Proactive Supply Chain Crisis Management},
  author       = {S. Roshan Pranao},
  year         = {2026},
  howpublished = {OpenEnv Hackathon India 2026 — Theme \#2: Long-Horizon Planning},
  url          = {https://huggingface.co/spaces/roshan5emerald/logiflow-rl}
}
```

---

*Submitted to the Meta × PyTorch × OpenEnv × Scaler Hackathon India 2026 — Theme #2: Long-Horizon Planning & Instruction Following*