Chris Paxton — Sharpa robot at IROS catches randomly falling foam cylinders with faster reaction time than the tweeter

TL;DR

At IROS 2026, Chris Paxton posts a video of a new Sharpa Robotics hand catching foam cylinders dropped from random positions — an in-person demo of reactive dexterous manipulation on the SharpaWave-lineage hardware. Paxton reports the robot outperformed his own attempt at the same task. No paper, no policy details, no metrics — the tweet is a first-person “demo happened” observation, notable as another public appearance of the Sharpa dexterous hand platform alongside the Origami Challenge dataset (Sharpa releases the Origami Dataset for the IROS 2026 Robotic Origami Challenge) and the ADEPT sim-to-real pipeline (ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning).

Key claims

A new Sharpa robot on the IROS 2026 exhibit floor catches foam cylinders dropped from random positions in real time [tweet body].
The tweeter’s own human-arm attempt at the same task was worse than the robot’s — offered as an informal reaction-time comparison, no measurements [tweet body].
No claims are made about the underlying policy, sensor stack, control frequency, or training data — this is a public-demo observation only [tweet body].

Method

Not disclosed. The tweet is a single video clip plus a one-line reaction; there is no method write-up, project page, or paper. The visible setup is a Sharpa dexterous hand mounted on an arm, catching foam cylinders released above it, apparently without a fixed release trajectory.

Results

Not applicable — no quantitative metrics. The tweet’s evidence is one video of successful catches and Paxton’s subjective observation that his own attempt was worse.

Why it’s interesting

Adds a second-hand third-party datapoint on Sharpa hardware in the wild: after the IROS 2026 Origami Challenge dataset release (Sharpa releases the Origami Dataset for the IROS 2026 Robotic Origami Challenge) and the ADEPT sim-to-real dexterity pipeline that trained on a 29-DoF Flexiv-Sharpa hand (ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning), this is the Sharpa platform showing up at IROS on a reactive task rather than a delicate contact-rich one — catching randomly falling objects requires perception + control loop latency the paper-folding and in-hand-rotation demos do not stress. Complements the Tactile sensing for manipulation thesis about fast tactile-reactive control obliquely — an unstructured falling object contacts the hand at unpredictable pose and time, exactly the regime where high-rate touch is supposed to help — but the demo predates any technical disclosure, so this file is a bookmark for whatever paper or blog post lands next, not a claim about the underlying system.