Event2Reward-Qwen3.5-4B-LIBERO-Spatial

The LIBERO-Spatial reward model of Event2Reward: Learning Robotic Rewards from Progress Relations (project page, code). Given a task description, the episode's start image, a goal image from an independent successful trajectory and five recent frames, it estimates task progress in [0, 1]; the change in this estimate is a dense reward for reinforcement learning.

Model

Qwen3.5-4B with a 101-bin progress head and a three-class trend head, in BF16. Fine-tuned from Stage2 on LIBERO-Spatial for 500 steps, with temporal supervision on successful demonstrations and progress relations on event-annotated rollouts. In the paper, residual SAC with this reward reaches 46.92 ± 6.94% policy success on LIBERO-Spatial (mean ± standard deviation over five seeds); this is a result of the full RL system, not of the checkpoint alone.

Use

Install the Event2Reward package, then:

from event2reward import Event2Reward, Observation

model = Event2Reward.from_pretrained("niejunnan25/Event2Reward-Qwen3.5-4B-LIBERO-Spatial")
observation = Observation(
    task="put the bowl on the plate",
    start="start.png",                           # first frame of the episode
    goal="goal.png",                             # last frame of another successful episode
    history=[f"h{i}.png" for i in range(5)],   # five recent frames at 10 Hz, oldest first
)
print(model.predict([observation]))             # [{'progress': ..., 'trend': ..., ...}]

Feed frames from one external camera, sampled at 10 Hz; early in an episode, repeat the first frame. The goal must never come from the episode being scored. See the inference guide for trajectories, batching and the command line, and the training guide for the LIBERO fine-tuning configurations.

Limitations

Progress is a learned estimate, not a success probability: it can decrease, underestimate a completed task or overestimate a failed one, and no threshold separates success from failure. The trend output is auxiliary and not a validated event detector.

License and citation

Apache-2.0, derived from Qwen3.5-4B; see LICENSE, NOTICE and LICENSES.

@misc{nie2026event2reward,
  title         = {Event2Reward: Learning Robotic Rewards from Progress Relations},
  author        = {Nie, Junnan and Fan, Zitian and Zhang, Jiachen and Lao, Junyi and Zhang, Likui and Liu, Chenghao and Lu, Xiaoyang and Wang, Yunbo and Zhang, Yixue and Zhai, Linbo and Li, Jiayi and Wang, Mengzhe and Wu, Xichang and Wang, Bin and Liu, Chang and Zhang, Tianle and Huang, Songfang},
  year          = {2026},
  eprint        = {2610.39212},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.39212}
}