ActiveArena: Benchmarking and Understanding Active Perception in Robotic Manipulation

Yibo Li^1,2^*, Enshen Zhou^1,2^*, Rui Chen¹, Yanjun Ding¹, Mengzhen Liu^2,3, Yi Han^1,2, Jiabo Zhan⁴, Lipeng Wang^1,2, Shanghang Zhang^2,3, Lu Sheng^1,2^†

¹ Beihang University ² Beijing Academy of Artificial Intelligence
³ Peking University ⁴ Tsinghua University

^* Equal contribution. ^† Corresponding authors.

▧ Paper ◫ Dataset ⌘ ActiveArena-Sim ⌘ ActiveArena-VLA ◉ Checkpoints

Move to see.
Remember to act.

ActiveArena is the first comprehensive and standardized simulator–benchmark–baseline suite for active perception in robotic manipulation.

Abstract

Active perception and manipulation are crucial for robots to interact with complex scenes. Existing benchmarks struggle to evaluate how robots effectively acquire and maintain information in memory in an active manner. To this end, we introduce ActiveArena-Sim, an active-perception simulator with controllable viewpoints and large-scale workspaces as the foundation. Built on this, we propose ActiveArena-Bench, which comprises 35 tasks across 5 fine-grained categories, covering visual exploration and interactive information acquisition.

Each task is difficult to solve from passive observations alone, requiring multi-round evidence acquisition and memory-based reasoning. The benchmark provides rich memory annotations, standardized training data, and ID/OOD protocols featuring disjoint scenes, unseen distractor configurations, and novel backgrounds. We further present ActiveArena-VLA, a modular suite of 13 vision-language-action configurations for controlled studies of memory writing, memory capacity, proprioceptive state, subtask supervision, and high-level planning.

Active Perception Rollouts

Representative ActiveArena rollout showing active viewpoint changes during manipulation — Representative rollouts show how the robot searches for visual evidence, updates its memory, and completes the manipulation task.

ActiveArena Benchmark

A benchmark for manipulation when the initial observation is insufficient. Policies must search for evidence, retain it across views, and use it to execute an 18-dimensional action.

ActiveArena simulator, task families, and benchmark protocol — Benchmark design, task families, controllable Astribot S1 embodiment, and the fixed ID/OOD evaluation protocol.

Modular VLA Suite

The release separates visual memory, language supervision, proprioceptive state, and action prediction so that each source of active-perception capability can be studied independently.

ActiveArena modular VLA framework — Modular vision-language-action configurations studied in ActiveArena. The released checkpoints use the continuous OFT action head.

Planner-guided memory management provides a sparse-memory alternative for active perception.

Rollout Demonstrations

The head camera is the policy observation. Observer and world cameras are included for visualization only.

DEMO / OBSERVER VIEW

Task

Scene split

Camera

Observer and world views are shown for explanation only. The policy receives the head camera view.

Simulation Results

Task-macro success rates reported in the paper. Each cell shows ID / OOD (%); Avg. is computed over all 35 tasks.

Method	Mem.	SS	SL	ML	MD	IA	Avg.
FAST-WAM	No	59.60 / 4.00	40.38 / 1.00	25.67 / 0.00	22.40 / 5.20	12.00 / 3.33	35.60 / 2.06
π_0.5	No	41.20 / 5.60	22.38 / 0.63	1.67 / 0.00	1.60 / 1.20	2.67 / 5.33	16.86 / 1.71
SaPaVe	No	48.80 / 40.80	19.88 / 5.88	10.67 / 3.00	6.00 / 2.80	25.33 / 12.67	20.91 / 10.51
HiF-VLA	Yes	16.80 / 14.00	6.88 / 1.13	0.00 / 0.00	17.60 / 17.20	30.67 / 18.67	10.69 / 6.57
MemER	Yes	81.20 / 66.00	33.88 / 18.63	18.33 / 5.67	11.60 / 6.40	40.00 / 42.00	35.31 / 23.43
MemoryVLA	Yes	46.00 / 34.40	20.88 / 6.50	14.00 / 2.00	9.20 / 4.40	28.00 / 27.33	22.23 / 11.20
ActiveArena-OFT	Yes	80.40 / 70.80	74.50 / 54.38	53.33 / 33.33	31.60 / 27.20	44.00 / 35.33	62.97 / 47.60
ActiveArena-Plan	Yes	79.20 / 69.60	73.38 / 52.63	49.67 / 30.00	28.00 / 24.80	42.00 / 33.33	60.97 / 45.54

BibTeX

@misc{li2026activearenabenchmarkingunderstandingactive,
      title={ActiveArena: Benchmarking and Understanding Active Perception in Robotic Manipulation},
      author={Yibo Li and Enshen Zhou and Rui Chen and Yanjun Ding and Mengzhen Liu and Yi Han and Jiabo Zhan and Lipeng Wang and Shanghang Zhang and Lu Sheng},
      year={2026},
      eprint={2609.24124},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2609.24124},
}