Skip to content

Grip — waste-sorting robots picking deformable trash with foundation models (YC S26)

Launch tweet for Grip (YC S26), a robotics startup applying foundation-model manipulation policies to waste sorting. The thesis is a two-part timing claim: foundation models can now grasp deformable, entangled, never-before-seen objects that legacy sorting machines punt on, and the hardware to run those policies has gotten cheap enough that per-pick economics finally close on trash. Grip installs modular robotic cells and pools data across the fleet — each pick trains the whole network. Initial deployment target is plastic contamination in organic-waste streams, with a stated roadmap of expanding to every waste stream where humans currently do the residual picking. Founded by ETH Zürich robotics engineers.

  • Foundation-model manipulation policies now handle deformable + entangled + open-vocabulary objects well enough to pick from real waste streams [tweet body].
  • Robot hardware cost has crossed the threshold where waste-sorting economics work [tweet body].
  • Grip operates a fleet-learning loop: every pick from every installed cell feeds back into the shared model [tweet body].
  • Beachhead is plastic contamination in organic waste; positioning as a general manipulation platform for waste infrastructure [tweet body].

Not disclosed in the tweet. From the framing — modular cells, fleet-wide learning signal per pick, open-vocabulary grasping of deformables — the implied stack is a VLA-style policy (or a grasp foundation model with language conditioning) driving commodity manipulator hardware, with centralized retraining on aggregated per-pick data (likely including failure telemetry). No details on backbone, action head, sensing modality, or gripper design.

None reported. This is a company-launch tweet, not a technical writeup — 10K views and short congratulatory replies at filing time.

An application-domain data point for the VLA Models cluster: the pitch reads as a bet that current-generation manipulation foundation models are already good enough for a specific real-world vertical (deformable waste), not a research claim that they will be. Concretely comparable to Gemini Robotics 1.5 brings AI agents into the physical world, whose worked example is context-grounded recycling-sorting via tool-calling — Grip’s fleet-learning angle is what Gemini Robotics doesn’t emphasize. It also fits the broader pattern tracked in Enact launch — post-training infrastructure that generates targeted recovery data for robotics VLAs (Enact) and MicroFactory — 99.9% reliability via $5 human-in-the-loop DAgger retraining on Jetson (Ilir Aliu × Igor Kulakov podcast) (MicroFactory): startups treating fleet-collected recovery data as the moat rather than the base policy.