End2Race

End2Race is an end-to-end policy and simulation environment for multi-vehicle autonomous racing on F1TENTH. It maps 2D LiDAR scans and vehicle speed directly to steering and speed commands at 40 Hz. This repository provides a quick demo of the environment and pretrained policy.

• Paper: End2Race: An End-to-End Learning Framework for Multi-Vehicle Autonomous Racing
• Code & Full Documentation: michigan-traffic-lab/End2Race

If you find this model helpful, please give us a star on GitHub!

Installation

pip install "https://huggingface.co/zhijieq/End2Race/resolve/main/end2race-1.0.0rc2-py3-none-any.whl"

Quickstart

Single-Vehicle Timed Trial

import gymnasium as gym
from transformers import AutoModel

import end2race  # registers End2Race-v0

policy = AutoModel.from_pretrained("zhijieq/End2Race", trust_remote_code=True)

env = gym.make(
    "End2Race-v0",
    track="Austin",                 # Austin, Hockenheim, MoscowRaceway, Nuerburgring
    mode="single",
    laps=1,                         # target laps to complete
    render_mode="human",            # "human" for interactive viewer or None
)

observation, info = env.reset()
policy.reset()
done = False

while not done:
    action = policy.act(observation)
    observation, reward, terminated, truncated, info = env.step(action)
    done = terminated or truncated

env.close()
print(f"Outcome: {info['outcome']} | Lap Times: {info['lap_times']} | Avg Speed: {info['avg_speed']:.2f} m/s")

Head-to-Head Racing

import gymnasium as gym
from transformers import AutoModel

import end2race  # registers End2Race-v0

policy = AutoModel.from_pretrained("zhijieq/End2Race", trust_remote_code=True)

env = gym.make(
    "End2Race-v0",
    track="Austin",                 # Austin, Hockenheim, MoscowRaceway, Nuerburgring
    mode="multi",
    opponent_raceline="raceline0",  # raceline0 (inner), raceline1 (center), or raceline2 (outer)
    speed_scale=0.6,                # opponent speed multiplier
    start_idx=0,                    # ego start waypoint index
    gap=15,                         # opponent lead in waypoints (~0.2 m each)
    duration=8.0,                   # episode duration in seconds
    render_mode="human",            # "human" for interactive viewer or None
)

observation, info = env.reset()
policy.reset()
done = False

while not done:
    action = policy.act(observation)
    observation, reward, terminated, truncated, info = env.step(action)
    done = terminated or truncated

env.close()
print(f"Outcome: {info['outcome']} | Avg Speed: {info['avg_speed']:.2f} m/s")

Environment Parameters

Interface

Observations (gymnasium.spaces.Dict)

• observation.lidar: float32[1440] range measurements over 360° (max 30 m).
• observation.speed: float32[1] longitudinal speed in m/s.
• observation.pose: float32[3] ego pose [x, y, heading] in world coordinates.

Actions

float32[2] as [steering_angle_rad, desired_speed_m_per_s]. Steering is clipped to [-0.4189, 0.4189] rad (±24°).

Policy Methods

• policy.reset(): Resets recurrent GRU hidden state. Call before each episode.
• policy.act(observation): Downsamples LiDAR to 180 beams, embeds speed, and computes [steering, speed].

Episode Termination

• terminated: Set on collision (both modes) or lap completion ("single" mode).
• truncated: Set when duration expires ("multi" mode).
• info["outcome"]: "overtake" or "follow" (head-to-head), "finished" (timed trial), or "collision".

Citation

@misc{qiao2025end2race,
      title={End2Race: An End-to-End Learning Framework for Multi-Vehicle Autonomous Racing},
      author={Zhijie Qiao and Haowei Li and Zhong Cao and Henry X. Liu},
      year={2025},
      eprint={2509.16894},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2509.16894},
}