ActiveScale: Scaling Active Perception for Robots across Model, Data, and Hardware
Shuai Zhou, Kaisheng Pang, Wenxuan Song, Wenjie Zhang, Xinhu Zheng, Haoang Li
cs.RO · cs.LG
Quick verdict
Pose tokens + 1000h mid-training lift π0.5 from 30% to 70% on five active-perception tasks; promising, but single baseline, n=20 trials, no significance tests.
Summary
ActiveScale extends the π0.5 VLA for active perception via three coordinated contributions. Model: front-camera history (current frame plus three frames at 16-frame intervals) with learnable camera tokens supervised by a VGGT-style head predicting 9-D translation/quaternion/FOV, enabling cross-view association. Data: a two-stage recipe—human–robot mid-training on over 1000 hours of egocentric human data (EgoLive, EgoVerse) balanced 1:1 with robot data (AgiBot World, RoboCOIN, AMP), followed by task-specific post-training. Hardware: AMP, a modified AgileX Cobot-Magic with a third arm carrying an actuated camera, teleoperated by one operator through a Meta Quest 2. On five real-world task families (Bag, Drawer, Table, Pot, Box) with 150 demonstrations each and 20 evaluation rollouts, ActiveScale improves mean success rate from 30.0% (post-trained π0.5) to 70.0% and task progress from 41.6% to 78.4%. Ablations show camera tokens are critical (history alone hurts on Bag/Pot) and mid-training adds ~8 points average SR. Main limitations: the headline comparison confounds architecture and mid-training changes against a single baseline; no active-perception VLA baselines (SaPaVe, ActiveMimic, EgoMI) are empirically compared; results rest on 20 rollouts per task with no variance estimates; mobile-manipulation evaluation is a single teleoperation demonstration rather than a learned policy.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Adding pose-grounded camera tokens to a π0.5-style VLA with front-camera history enables cross-view reasoning and substantially improves active-perception manipulation success.
- A 1000-hour human–robot mid-training stage on egocentric and multi-embodiment robot data is the first active-perception training effort at this scale (>5× prior work) and improves downstream success by ~8 points on average.
- The AMP platform allows a single operator to teleoperate viewpoint, bimanual manipulation, and mobile-base motion simultaneously, enabling scalable coordinated demonstration collection.
- ActiveScale raises mean success rate from 30.0% to 70.0% and mean task progress from 41.6% to 78.4% versus post-trained π0.5 across five real-world task families.
Strengths
Well-isolated ablation of the core modeling idea
Sec. IV-C2 compares π0.5, history-only, and history+camera-token variants: history alone is task-dependent (Table SR 10%→50% helps, but Bag SR 40%→20% hurts), while adding pose-grounded camera tokens lifts all five families (e.g., Bag 20%→50%, Pot 10%→50% vs history-only). This honestly shows raw history is insufficient and pins the gain on pose grounding rather than extra input.
Large-scale, carefully composed mid-training data
Sec. III-D1 details a 1:1 human:robot mix over EgoLive/EgoVerse and AgiBot World/RoboCOIN/AMP (~1000h total, claimed 5× EgoMI's ~200h), with sensible choices: camera-frame re-expression of human trajectories, per-source normalization, masked unavailable action/pose dimensions, and weighted losses (λH=λC=0.2).
Complete, reproducibly specified hardware/teleop design
Strict, staged evaluation criteria
Deployment practicality addressed
Weaknesses
Headline comparison conflates two changes against one baseline
No empirical comparison to prior active-perception methods
No statistical treatment of small evaluation samples
Key training and data details are underspecified
Mobile manipulation is demonstrated, not evaluated
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ActiveScale and how does it improve robot active perception?
ActiveScale is a framework built on the π0.5 vision-language-action model that lets a robot actively move its camera to gather task-relevant views during manipulation. It adds a history of front-camera frames with learnable camera tokens supervised to predict camera pose and field of view, mid-trains on about 1000 hours of egocentric human video plus multi-robot data, and uses a new teleoperation platform (AMP) with an independently actuated camera arm. On five real-world occlusion and search tasks it raises average success rate from 30% to 70% versus a post-trained π0.5 baseline.
What is the AMP platform in the ActiveScale paper?
AMP (Active-perception Mobile-manipulation Platform) is a modified AgileX Cobot-Magic: a mobile base with two 6-DoF Piper arms for bimanual manipulation plus a third arm carrying an Orbbec camera as an actively controlled 'eye.' A single operator wearing a Meta Quest 2 moves the camera with their head and the arms with handheld controllers, and drives the base with controller buttons, so viewpoint changes, manipulation, and navigation are recorded together at 30 Hz as training demonstrations.
Why does the paper train on egocentric human videos like EgoVerse and EgoLive?
Human first-person video naturally contains coordinated head-camera motion—people reposition their viewpoint to search for objects and resolve occlusions—so it offers a scalable source of active-perception behavior without robot hardware. ActiveScale mixes this human data 1:1 with robot demonstrations during a mid-training stage so the model acquires viewpoint-control priors from humans while keeping robot-aligned action generation, which their ablation shows adds about 8 points of downstream success on average.
Is ActiveScale's code, data, or the AMP platform publicly available?
The paper points to a project website (active-scale.github.io) for materials, but the extracted text states no public code release, does not commit to releasing the AMP teleoperation datasets, and does not release model checkpoints. Replication currently requires rebuilding the three-arm platform and re-collecting the 150 demonstrations per task.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 4-8 were available only as extracted numeric text; qualitative behavior claims (e.g., 'stalls without initiating a grasp') and figure quality could not be visually verified.
- The supplementary video and project website content were not accessible, so the mobile-manipulation demonstration and data-release statements rest on the paper text alone.
- Several cited works are 2026 arXiv preprints (e.g., SaPaVe, ActiveMimic, EgoLive, EgoVerse); their reported numbers could not be cross-checked within this review.
- Hyperparameters (learning rates, batch sizes, training steps) do not appear in the extracted text; they may exist in an appendix not included in the extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.