Questions robotics teams actually ask us.

Yes. Neuracore is built for both. Most teams start with imitation learning, training ACT, Diffusion Policy, and π0 directly on teleoperated demonstrations, and then use reinforcement learning to fine-tune a policy in simulation or on hardware where demonstrations alone plateau. You can run either independently, or chain them: pre-train with imitation learning, then improve with RL against a reward you define, all tracked as versioned runs on the same dataset.

We support the common manipulation policy families out of the box: ACT, Diffusion Policy, π0, and vision-language-action (VLA) models that take camera frames plus a language instruction and output actions. You can pick an architecture per run, bring your own model code, or fine-tune an open VLA checkpoint on your data. Training runs are versioned so you can compare success rates across architectures on the same held-out evaluation set.

Yes. The data schema and policies are not hard-coded to a single 6-DoF arm with a parallel gripper. We handle bimanual setups, high-DoF dexterous hands, mobile bases, and mixed action spaces. Demonstrations record the full observation and action vectors for whatever embodiment you use, and policies are configured to match that action space.

Neuracore ingests demonstrations from common teleoperation rigs, leader-follower arms, VR controllers, space mice, and teleoperation gloves, as well as kinesthetic teaching and scripted trajectories. If your rig can stream synchronized observations and actions, we provide an SDK to log them in the platform format. Timestamps, camera streams, and proprioception are aligned automatically.

Yes. We provide ROS 2 packages for both ends of the loop: recording demonstrations from live topics during teleoperation, and running a trained policy as a node that subscribes to observations and publishes actions. If you are on ROS 1 or a custom middleware, the same data can be bridged through our SDK.

Yes. Trained policies export to a runtime you can run on-robot or on a nearby edge box, so control does not depend on a round-trip to the cloud. We support GPU and common edge accelerators, and report the inference latency and control frequency a policy achieves on your target hardware so you know it meets your loop rate before deployment.

Both. Our cloud-hosted platform is the fastest way to start, and if you already have a teleop rig you can begin for free. For teams with data-residency or air-gap requirements, Enterprise runs privately-hosted: the full pipeline, collection, curation, training, and the deployment runtime, inside your own VPC or on-prem cluster. The platform and workflows are identical; only where it runs changes.

We integrate with the common manipulation simulators (Isaac Sim / Isaac Lab, MuJoCo, and PyBullet-style environments) for generating synthetic demonstrations and running reinforcement learning rollouts. Simulated and real demonstrations live in the same dataset, so you can blend them in a training run and track exactly what mix produced a given policy.

Deployed policies stream success/failure outcomes and observations back to the platform. You see success rates over time, can trace a failure back to the exact states it struggled with, and flag those episodes for re-teaching, the failures become new demonstrations that feed the next training run.

Yes to both. You are never locked in. You can bring custom policy architectures and training code and run them on the platform, and you can export your full dataset, demonstrations, annotations, and versions, in open formats at any time. Your data and your models are yours.

You own all of your demonstration data, datasets, and trained policies. Data is encrypted in transit and at rest, access is scoped per project and role, and self-hosted deployments keep everything inside your own infrastructure. We do not train shared models on your data.

There are two ways in. If you already have a teleoperation rig, start free on our cloud-hosted platform: collect, curate, train, and export your data at no cost. When you need to scale, Enterprise adds full platform access, support, and security controls, cloud-hosted or privately-hosted in your own infrastructure, priced to your setup rather than a fixed per-robot fee. Talk to us for an Enterprise quote.

Yes. Enterprise includes an optional Pilot Evaluation: our forward-deployed engineers come and stand up your teleoperation system with you, then collect, train, and deploy one or two real skills measured against an agreed success target. The pilot fee covers that engineering time and your first three months of platform subscription, so you prove the workflow on your own hardware before a broader rollout.