PortSimEnv v1

Re-plan container-ship dockings at the Port of Barcelona. Each task is a real week (or two or three) at one container quay, built from the port's own 2024 records, with disruptions added: late and bunched ships, closed quay sections, crane breakdowns, gales under the port's wind rules, traffic diverted from the other terminal, emergencies and priority cargo. The agent decides when and where each ship docks and with how many cranes, and is graded once, deterministically, against the plan a CP-SAT solver proved optimal.

Environment (OpenEnv Space, 3D viewer, eval explorer)	FineEnvs/PortSimEnv
Article	Simulation RL Environments
Bucket (3D twin data, raw eval rollouts)	FineEnvs/PortSimEnv
Code	adithya-s-k/FineEnvs: 07-simulation-environments/portsim-v1
Ideas for v2, v3, post-training, data	GitHub Discussions

Configs

config	split	rows	what
tasks	train	1,050	training tasks
tasks	eval	50	held-out tasks: whole week groups that never appear in train
rollouts	eval	300	6 models x 50 eval tasks: transcript, tool calls, final plan, grade
calls	train	1,784	the 2024 container calls at quays 36A (BEST) and 24B (APM Terminals) the tasks are built from

tasks: system_prompt and situation are exactly what the environment sends (the rules, then the opening message, which is also what get_situation() returns). task is the full task as JSON (ships, closures, disruptions, rules). optimal_plan / optimal_cost are the CP-SAT reference (proven_optimal says whether optimality was proven); naive_plan / naive_cost re-plan by pushing ships later. A plan is a JSON list of {"ship": id, "berth_hour": h, "section": s, "cranes": c}.

rollouts: one row per (model, task) from the eval run dock-eval50: 12 turns and 32k output tokens per turn, the same three tools. messages is the full transcript, steps the tool calls with each check_plan result, grade the final grade.

Results on the 50 eval tasks

model	mean reward	standard	busy	storm	extreme	submitted	valid	optimal
GPT-6.1 Sol	0.888	1.00	0.90	0.86	0.82	50	50	26
Claude Sonnet 5.5	0.782	0.92	0.80	0.68	0.77	49	49	10
GLM-5.3-Flash	0.470	0.77	0.50	0.46	0.24	36	25	9
Qwen3.8-2.4T	0.380	0.58	0.45	0.33	0.21	34	23	5
GLM-5.3	0.313	0.62	0.34	0.22	0.15	17	16	10
Qwen3.8-27B	0.211	0.40	0.27	0.22	0.00	16	13	3

Reward

Deterministic, no LLM judge. A plan that breaks any rule scores at most 0.2 (0.2 x the share of ships placed cleanly). A valid plan's cost (hours each ship leaves after its due time x its size x its priority, plus penalties for late emergency dockings and moved ships) is compared with the optimum: gap = (cost - optimum) / (optimum - unavoidable + 100), reward = 0.2 + 0.8 * exp(-gap / 0.5). No submission scores 0.

Use

import json
from datasets import load_dataset

tasks = load_dataset("FineEnvs/PortSimEnv", "tasks", split="eval")
task = tasks[0]
print(task["system_prompt"], task["situation"], sep="\n\n")

Every task can be played on the environment Space (OpenEnv; MCP tools get_situation, check_plan, submit_plan). Submitting the reference optimum scores 1.0:

from openenv.core.env_server.mcp_types import CallToolAction
from openenv.core.mcp_client import MCPToolClient

env = MCPToolClient("https://fineenvs-portsimenv.hf.space").sync()
env.reset(task_id=task["task_id"])
step = env.step(CallToolAction(tool_name="submit_plan", arguments={"plan": json.loads(task["optimal_plan"])}))
print(step.reward)  # 1.0

Source and licence

Contains data from the Port de Barcelona open data portal (https://opendata.portdebarcelona.cat/), via the 2024 snapshot in alberto-santini/berth-allocation-problems, licensed CC BY-SA 4.0. This dataset is shared under the same licence. Crane fleets, wind rules and handling rates come from the terminals' and the port's published information; disruptions are generated. Work in progress (v1).