One model, many hands
A single policy generalizes across kinematically different grippers and hands.
Dextra · GripSim
Dextra
Cross-embodiment VLA manipulation foundation model + fine-tuning toolkit.
Overview
What Dextra does
One vision-language-action model, pretrained across many gripper and hand shapes, with an embodiment-conditioned fine-tuning kit so a new hand inherits skill from a small dataset.
- One model, many hands
- Fine-tuning toolkit
- Morphology transfer
- Built-in eval harness
Key features
Inside Dextra
Built to be GPU-essential, NVIDIA-aligned, and production-ready.
Fine-tuning toolkit
Point it at your hand and cell, fine-tune on a small dataset, ship a working policy.
Morphology transfer
Adapters let a new gripper inherit capability with minimal new data.
Built-in eval harness
Benchmark grasp success and cycle time against open and hand-coded baselines.
Closed-loop ready
Pairs with Tactyl's contact predictions for reactive, in-hand control.
Improves with use
Every deployment feeds the shared policy through the GripSim data flywheel.
Specs
Technical specifications
| Model type | Vision-language-action (VLA) |
|---|---|
| Training | Imitation + RL across morphologies |
| Compute | NVIDIA DGX / HGX (Blackwell) |
| Inference handoff | TensorRT → Reflex runtime |
| Delivery | Pretrained policy + fine-tuning toolkit |
| Sensors | Vision, optional tactile via Tactyl |
Gallery
Dextra in action
Use cases
Where Dextra fits
- Warehouse bin-picking of mixed, novel items
- Machine tending across part variants
- Humanoid hand manipulation
- Fine assembly and kitting
Pricing
Ways to deploy Dextra
Start free. Scale to a fleet. Pricing shown is indicative, pre-launch.
Pilot
Project
- Scoped pilot
- Engineer support
- Benchmark report
Scale
Usage
- Production license
- Fleet updates
- Priority support
Enterprise
Custom
- Custom SLAs
- On-site support
- Roadmap input
FAQ
Dextra questions
Does it require NVIDIA hardware?
Yes — Dextra is GPU-essential; see the specs above for the exact stack.
Can I use it standalone?
Yes, Dextra works on its own and integrates with the rest of the GripSim stack.
How do I get started?
Talk to our team for a scoped pilot tailored to your robot and task.
Related
Explore the rest of the stack
Tactyl
Tactile-and-vision contact world model that predicts grasp outcomes.
Explore Tactyl →
Reflex
On-robot edge runtime for low-latency closed-loop manipulation.
Explore Reflex →
Simdex
Domain-randomized manipulation simulation + data-generation engine.
Explore Simdex →Why GripSim
Built for production robotics
The manipulation layer teams ship with, not a research demo.
Cross-embodiment by design
One policy runs across many grippers and hands — no per-robot rewrite.
Closed-loop with touch
Tactyl predicts slip and force so grasps react before they fail.
Runs on the robot
Reflex executes on Jetson at control-loop latency, with safety limits.
A moat that compounds
Every deployment feeds the shared model through the data flywheel.
Stay in the loop
Manipulation research, in your inbox
Benchmarks, sim-to-real notes, and product updates. No hype, no spam.
Put Dextra to work
Tell us your robot and task — we'll scope a pilot.