Short video: https://youtu.be/iIdum8FrMIw

Website: https://generaltrajectory.com

What we’re working on: Humanoid robots can do backflips and kung fu, but still struggle with economically useful work. Dexterous manipulation (human-level grasping and handling of everyday, novel objects in messy scenes) is the bottleneck to useful physical labor.

Why this matters: A lot of “robot demos” happen in tightly-controlled settings. Once you move into real environments, the distribution shifts immediately: clutter, lighting changes, partial occlusions, novel objects, weird geometries, slippery materials, etc. For many of these “difficult” objects, prior SoTA systems can collapse to 0% success.

Our approach: We combine a small, efficient set of human demonstrations with reward-guided training to improve the base policy’s grasps.

Concretely:

- Collect an efficient dataset of demonstrations (teleop)

- Train a reward model

- Use that reward to improve grasping behavior beyond pure imitation

Results so far: We see strong generalization to novel objects and scenes, including cases where strong baselines hit 0%. On the hardest objects, we get up to 63% gains, while maintaining near-perfect performance on standard objects.

Full details (technical write-up): https://www.generaltrajectory.com/technical-update

Deployment: We use our policy as an action expert inside a Vision-Language Action (VLA) model: a 2B VLM (Qwen3-VL) handles high-level reasoning/planning (“what/when”), and our 12-DoF dexterous policy executes the “how” (grasp/manipulate) on unseen objects.

We show autonomous real-world rollouts for:

- Box packing + folding (end-to-end with unseen items)

- Delivery fulfillment (picking unseen food items into DoorDash delivery bags)

We also open-sourced our teleop stack: https://github.com/GeneralTrajectory/dex-teleop

It uses vision + wrist trackers instead of data gloves → about $500 in hardware vs. ~$5,000.

The physical world represents 80% of global GDP. We believe this research represents progress toward a future in which AI systems are not limited to computer work and can contribute directly to the physical world.

If you’re interested in AI for the physical world (scientific R&D automation, defense, etc.), I’d love to hear your thoughts!