Research & Results
Our headline result — crowdsourced simulation data trains stronger policies — plus the published research the platform builds on.
Arctivar Robotics is built on published research and validated by our own experiments. The core team has led and contributed to foundational work at top-tier robotics and AI venues — including RSS, CoRL, CVPR, ICCV, and ICLR — and the platform builds on the canonical literature of interactive imitation learning. This page leads with our headline result, then the literature underneath.
Noisy Crowdsourcing Trains Strong Robots
The central question behind Arctivar: can large-scale, noisy, crowd-sourced simulation data serve as a sufficient substrate for embodied pretraining? Our answer is a qualified yes — dirty at the trajectory level, clean in the aggregate, provided diversity and coverage clear a critical bar.
We continually pretrain Physical Intelligence's π0.5 on the Arctivar dataset, fine-tune on LIBERO with an identical recipe across conditions, and evaluate on LIBERO-Plus, a perturbation benchmark that stress-tests a policy across camera, lighting, sensor noise, background, layout, language, and robot variations. Pretraining on the full Arctivar set lifts overall success rate from 83.9 to 88.8 — a 5.8% relative improvement — and beats a volume-matched RoboCasa365 control by 31.3 points.
| Pretraining | Demos | Overall | Camera | Light | Noise | B.G. | Layout | Language | Robot |
|---|---|---|---|---|---|---|---|---|---|
| π0.5 (vanilla) | 0 | 83.9 | 72.5 | 98.2 | 82.5 | 94.4 | 82.9 | 89.6 | 74.4 |
| + ARCTIVAR-25% | 0.25 N | 84.7 | 77.5 | 91.2 | 86.2 | 100.0 | 85.5 | 81.8 | 76.9 |
| + ARCTIVAR-50% | 0.50 N | 85.7 | 68.8 | 98.2 | 91.2 | 96.3 | 88.2 | 84.4 | 79.5 |
| + ARCTIVAR-100% (ours) | N | 88.8 | 83.8 | 96.5 | 96.2 | 98.1 | 85.5 | 88.3 | 78.2 |
| + RoboCasa-matched (control) | N | 57.5 | 35.2 | 79.5 | 63.2 | 81.7 | 68.0 | 49.2 | 39.4 |
LIBERO-Plus success rate (%). N = the full Arctivar demonstration set; the control uses the same trajectory count drawn from RoboCasa365.
Two things stand out. First, overall performance scales with data — 84.7 → 85.7 → 88.8 across 25%, 50%, and 100% of the Arctivar set. Second, the control is decisive: RoboCasa-matched, at the identical trajectory count, lands at 57.5 — far below vanilla. Same volume, much worse result. The gains come from the Arctivar pipeline — crowdsourced tasks, diverse demonstrations, and large-scale randomization — not from data volume alone.
Where the Gains Land
Broken out by perturbation axis, the largest improvements fall where the augmentation is aimed — sensor noise and camera:
| Arctivar | vs. vanilla π0.5 | vs. RoboCasa-matched |
|---|---|---|
| Sensor Noise | +13.7 | +33.0 |
| Camera | +11.3 | +48.6 |
| Robot | +3.8 | +38.8 |
| Background | +3.7 | +16.4 |
| Layout | +2.6 | +17.5 |
| Language | −1.3 | +39.1 |
| Light | −1.7 | +17.0 |
| Overall | +4.9 | +31.3 |
Absolute gain in points, ARCTIVAR-100% against vanilla π0.5 and against the volume-matched RoboCasa365 control.
Lighting and language sit within noise of vanilla — the augmentation does not target them — while every axis clears the matched-volume control by a wide margin. That pattern is the point: Arctivar data buys robustness where distribution shift actually bites, and it does so because of what the data contains rather than how much of it there is.
What This Changes
Egocentric data already made this case for perception and world models. These results extend it to control: individual trajectories can be imperfect as long as the distribution they form is diverse and well covered, because what a policy learns from is the distribution, not any single demonstration.
So the unit of quality control moves. Stop grading the trajectory — grade the distribution, then let closed-loop refinement finish the last mile: the policy's own failures reopen the tasks it is weakest at, and human-gated correction supplies exactly the data the distribution is missing.
The Dataset Behind the Result
The Arctivar dataset used in this experiment spans 207 manipulation tasks, 50,000+ human demonstrations, and 60,000+ scene variants across 7 scene categories (kitchen, toys, personal care, office, household, tools, tabletop). Each task carries a language instruction, a parameterized scene, its assets, and a structured success checker; each trajectory carries synchronized RGB-D from two camera views, robot and object states, actions, metadata, and success labels. Every trajectory passes format verification, physical-consistency and success checking, smoothing, and resampling before IsaacSim rendering — the same Data Pipeline that runs in production. Refinement cuts mean acceleration by 63.9% and mean jerk by 80.8% against raw teleoperation. Full methodology, ablations, and the evaluation protocol are in the paper below.
The Arctivar Paper
ARCTIVAR: A Growable Community-Driven Data Engine for Scalable Robot Manipulation is the peer-facing account of the system and the results above — browser-based teleoperation, automatic task generation and validation, the refinement and augmentation pipeline, the versioned task snapshots, and the held-out evaluation protocol used to compare policies.
Core Papers
RoboVerse: Towards a Unified Platform for Scalable and Generalizable Robot Learning
The unified cross-simulator framework that powers Arctivar's simulation engine. RoboVerse integrates multiple simulation environments (MuJoCo, IsaacSim, and others) into a single pipeline with consistent APIs, data formats, and asset specifications — enabling trajectories to flow seamlessly between lightweight collection frontends and high-fidelity augmentation backends.
GraspVLA: Generalizable Grasping with Vision-Language-Action Models
Demonstrates that models trained purely on simulation data can achieve high zero-shot success rates on real robots — validating the core Arctivar thesis that high-quality simulation data, properly augmented, is sufficient for real-world generalization.
SkillBlender: Towards Versatile Humanoid Whole-Body Loco-Manipulation via Skill Blending
Fuses atomic robotic skills (grasp, move, place) into complex, multi-step behaviors — bridging basic motion primitives and high-level task reasoning across diverse embodiments including humanoids.
VGDP: Visual-Geometry Diffusion Policy for Robust Generalization
Uses modality-wise dropout to enforce balanced use of RGB and point-cloud cues, achieving 39.1% average improvement over baselines across 18 simulated and 4 real-world tasks — with particular strength under visual and spatial perturbations.
Interactive Imitation Learning & Post-training
The foundations of the platform's closed-loop data engine — why demonstrations alone hit a ceiling, and how human-gated correction removes it. See Model Training for how these methods run on Arctivar.
| Paper | Venue | Relevance | Link |
|---|---|---|---|
| Efficient Reductions for Imitation Learning (Ross & Bagnell) | AISTATS 2010 | Covariate shift: behavior cloning's compounding errors grow quadratically with the task horizon | PMLR |
| A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning (DAgger; Ross, Gordon & Bagnell) | AISTATS 2011 | Dataset Aggregation: expert labels on the learner's induced state distribution reduce error growth to linear | 1011.0686 |
| HG-DAgger: Interactive Imitation Learning with Human Experts (Kelly et al.) | ICRA 2019 | The human decides when to take over and when to hand back — the gating scheme behind post-training tasks | 1810.02890 |
| SOP: A Scalable Online Post-Training System for Vision-Language-Action Models (AgiBot) | arXiv preprint, 2026 | Scalable online post-training: deployed policies keep improving from on-policy experience and human interventions | 2601.03044 |
| Automated Creation of Digital Cousins for Robust Policy Learning (Dai et al.) | CoRL 2024 | Digital cousins: varied scene instances with similar affordances, in the spirit of Arctivar task variation | 2410.07408 |
Related Work
Papers by core team members and collaborators providing additional foundations in dexterous manipulation, object understanding, and generalizable robotic control.
| Paper | Venue | Relevance | Link |
|---|---|---|---|
| GAPartNet | CVPR 2023 Highlight | Cross-category object manipulation via generalizable part-level affordances | 2211.05272 |
| UniDexGrasp++ | ICCV 2023 Oral, Best Paper Finalist | Geometry-aware curriculum for universal dexterous grasping | 2304.00464 |
| RAM | CoRL 2024 Oral | Retrieval-based affordance transfer for zero-shot manipulation | 2407.04689 |
| SAGE | RSS 2024 Oral, Best Paper at SemRob | Semantic + actionable part reasoning for articulated objects | 2312.01307 |
| Rodrigues Network | ICLR 2026 Oral | Geometric SE(3) action representation for policy learning | 2506.02618 |
| FastTD3 | — | Streamlined RL for humanoid locomotion and control | 2505.22642 |
| DexGraspNet 2.0 | CoRL 2024 | Generative dexterous grasping in cluttered synthetic scenes | 2410.23004 |
Next: Solutions →