Model Training
Training robot policies on augmented simulation data, post-training them on human corrections, and validating sim-to-real transfer.
The final stage of the pipeline: turning collected data into robot policies that work in the real world. Training proceeds in two phases — pre-training on demonstrations, and post-training on human corrections — mirroring the two datasets the platform produces.
What We Train
Two classes of models are trained on the curated and augmented datasets:
Task-specific policies. Behavior cloning (BC) and diffusion policy (DP) models trained on individual task families (e.g., pick-and-place, drawer opening, pouring, button pressing). These are focused models optimized for specific manipulation skills, spanning rigid-object manipulation, articulated-object interaction, and contact-rich tasks requiring precision and temporal coordination.
General-purpose models. Models conditioned on language goals and visual context. Given an instruction like "pick up the red cube and place it on the plate" and a camera feed, the model predicts the robot's next action. These models aim for zero-shot generalization across tasks they haven't been explicitly trained on.
Post-training: Learning from Human Corrections
Why demonstrations alone are not enough
A policy trained by behavior cloning treats imitation as supervised learning — but the policy's own mistakes take it into states absent from the training distribution. This is the covariate shift problem (Ross & Bagnell, 2010): because the policy only ever saw states an expert visited, it receives no supervision on the states it produces once it acts on its own. Small errors compound, and the theoretical bound reflects it — behavior cloning's error grows quadratically with the task horizon. By construction, the situations in which a policy is weakest are the ones least represented in curated demonstrations.
Reducing this mismatch requires data drawn from the policy's own induced state distribution: on-policy corrective data, not more off-policy demonstrations.
DAgger, human-gated
DAgger (Dataset Aggregation; Ross, Gordon & Bagnell, 2011) attacks this directly: roll out the current policy, collect expert corrections on the states the policy actually visits, aggregate them into the training set, retrain, repeat. The training distribution converges toward the policy's induced state distribution, and the compounding-error bound drops from quadratic to linear in the horizon. Because the visited states include the policy's own mistakes, DAgger teaches recovery, not just imitation of error-free trajectories.
Arctivar uses the human-gated variant, HG-DAgger (Kelly et al., 2019), in which the human — not the algorithm — decides when to take control and when to hand it back. Human-gated correction is highly effective and notoriously expensive: it has traditionally required one human continuously supervising one robot. This is the constraint Arctivar removes. Run the loop in simulation, distribute the supervision across a community, and human-gated correction scales.
Scalable online post-training
Recent robot-learning work calls this paradigm scalable online post-training (SOP; AgiBot, 2026): rather than freezing a policy after offline training, the policy is deployed and keeps improving from the experience it generates in operation. SOP instantiates it with fleets of physical robots streaming trajectories and human interventions to a centralized learner. A consistent finding in this line of work: once a base policy exists, the marginal utility of on-policy correction exceeds that of additional offline demonstrations — demonstrations show diminishing returns, while a comparatively small volume of corrective data produces substantial improvement.
Arctivar realizes the same paradigm with different resources:
| SOP (robot fleets) | Arctivar | |
|---|---|---|
| Source of on-policy experience | Fleet of physical robots | Global community of contributors |
| Operating environment | The physical world | Simulation (browser + IsaacSim) |
| Corrective signal | On-site human operators | Distributed human-gated takeovers |
Running many contributors in parallel broadens coverage of the policy's induced state-action distribution, while centralized aggregation and retraining keep every contributor's corrections aligned with the current policy.
Operationalized as Post-training tasks
On the platform, this loop runs as post-training tasks: seeded, reproducible policy rollouts that contributors supervise and correct in the browser. Rollout initialization is designed so contributor effort goes toward correcting the policy, not re-specifying behavior it already executes reliably. As corrections accumulate, the policy is retrained; the stronger policy reaches new states it cannot yet handle, and those define the next round of targeted post-training tasks.
Training Protocol
The pre-training pipeline compares three data conditions to isolate the contribution of each processing stage:
- Raw teleoperation data — directly from the browser, no processing
- Cleaned data — after filtering, smoothing, and upsampling
- Cleaned + augmented data — after IsaacSim replay with domain randomization
All variants use the same task definitions and fine-tuning budget (typically 40–60 demonstrations per task), ensuring fair comparison and allowing the effect of data curation and realistic augmentation to be measured independently. Post-training then iterates on the strongest variant using the correction dataset.
Evaluation
In Simulation
Trained policies are evaluated under randomized initial conditions generated by the task pipeline:
- Success rate across varied initial object poses and configurations
- Robustness under domain randomization (unseen textures, lighting, physics)
- Generalization across object sets and scene layouts not seen during training
- Intervention rate — for post-trained policies, how often (and for how long) human takeover is needed across community rollouts; every post-training session doubles as an evaluation run
On Real Robots
Sim-to-real transfer is evaluated on a Franka Research 3 (FR3) robotic arm with its default end-effector — and it remains an open problem rather than a settled result. The deployment loop includes hardware adaptation, sensor calibration, and real-world task setup matching the simulation scenarios. The evaluation covers representative tabletop manipulation tasks: pick-and-place, drawer opening/closing, pouring, and button pressing.
Reported metrics include:
- Execution success rate
- Trajectory quality
- Generalization across physical conditions
Key Results

Continual pretraining of Physical Intelligence's π0.5 on the Arctivar dataset lifts overall success rate on the LIBERO-Plus perturbation benchmark from 83.9 to 88.8 — and a matched-volume control drawn from RoboCasa365 lands at 57.5, far below vanilla, confirming the gains come from the Arctivar pipeline, not data volume alone. Full results table and per-axis breakdown are in Research & Results.
More broadly, research by the core team has shown that models trained purely on simulation and synthetic data can reach high zero-shot success rates in real-world grasping — evidence for the thesis that simulation data, properly collected, cleaned, and augmented, carries policies toward reality. Closing that gap in general remains open work.
The academic foundations behind these results:
- GraspVLA (arXiv:2505.03233) — proves the efficacy of pure simulation and synthetic data for real-world grasping, achieving high sim-to-real transfer.
- SkillBlender (arXiv:2506.09366) — enables seamless fusion of atomic robotic skills into complex, multi-step behaviors across diverse embodiments including humanoids.
Known Limitations
Cross-simulator replay fidelity. Differences in contact models, actuator dynamics, and solver settings between MuJoCo and IsaacSim can introduce deviations, especially for contact-rich manipulation. Careful frame alignment and standardized conventions mitigate this, but replay should be viewed as a scalable augmentation mechanism rather than exact reproduction. Future work includes stronger tracking controllers and simulator calibration.
Labeling and task success signals. Teleoperation data lacks dense supervision such as rewards or temporally localized annotations. The platform combines task-specific success checkers with replay-time verification, and post-training adds a complementary signal: human takeover points mark precisely where the policy leaves a reliable regime. Denser automatic labeling remains an open direction.
Cross-embodiment transfer. The unified interface provides infrastructure-level compatibility, but robots with substantially different kinematics, dexterity, and control properties still require embodiment-specific adaptation. Addressing this limitation may require embodiment-aware retargeting, canonical action spaces, or learned motion translation.
Next: Research →