Flip Anything with Agentic Robotics Self-Learning

Shuangyu Xie, Kaiyuan Chen, Ken Goldberg

November 2026

Real-world trial: the bimanual YAM arms flip a clamp not seen during self-learning.

TL;DR A coding agent writes and refines a graph-based bimanual flipping policy purely by trial in simulation, with no demonstrations. Over self-learning iterations success on 24 objects rises from 24.7% to 95.0%, and the code runs zero-shot on real YAM arms.

What they do

Flipping an object 180 degrees on a table needs different strategies for different shapes: a hammer can be handed between two grippers, a wide multimeter has to be pushed and toppled. FlipAnything is a short benchmark and case study in which a coding agent (Claude Fable 5.1 in the Graph-as-Policy harness) writes one program that flips 24 objects with bimanual YAM arms, improving it only through simulated trials, without demonstration data.

How it works

Inputs. An English task description, CAD models of the 24 objects and the arm URDF. Simulation uses RoboSimStudio and Newton with a MuJoCo-based solver and CoACD convex decomposition.
Self-learning loop. The agent starts with a handover-only graph, evaluates it on all objects at randomized positions, reads success rates, traces and failure modes (infeasible grasps, lost contact, failed handovers, wrong orientation estimates), then revises perception, grasp selection, primitives, strategy selection and recovery.
Learned artifact. A graph that converts to executable code driven by RGB-D and proprioception. It picks between handover, repositioning, stand-and-topple and pushing, with closed-loop recovery such as grip verification and extra pushes.

Results

Over 360 episodes per iteration (24 objects, 15 each), success goes from 24.7% (89/360) with handover only to 80.6% (290/360) after adding toppling and pushing, and to 95.0% (342/360) for the best controller.
Per object: claw hammer 14/15 by in-air handover, both lock nuts and the multimeter 15/15, tape measure 14/15, pipe wrench 13/15, surgical scissors from 0/15 to 13/15.
Real robot, zero-shot on three unseen objects: clamp 2/3, blue tape 3/3, foldable helix screwdriver 1/3.

Limitations

A short paper that reports no baselines. The real-world test is three objects with three trials each, and the authors leave generalization to unseen objects to future work. The 95.0% is measured on the same 24 objects the code was refined on.

Videos

Real-world trial: flipping a roll of blue tape.

Simulation results on the 24 benchmark objects as the policy is refined across versions.

Figures

From Figure 1: The FlipAnything benchmark of 24 objects, the success curve over self-learning iterations (24.7% to 95... — From Figure 1: The FlipAnything benchmark of 24 objects, the success curve over self-learning iterations (24.7% to 95.0%) and real-world deployment of the flip skill.

From Figure 3: The generated Graph-as-Policy graph routes each observation through plan selection to the strategy com... — From Figure 3: The generated Graph-as-Policy graph routes each observation through plan selection to the strategy components, then re-observes and updates its belief. Blue nodes call the camera or robot; green nodes are agent-written scripts.

From Figure 4: Grasp selection for all 24 objects, drawn on a top-down view of the perceived height map, darker cells... — From Figure 4: Grasp selection for all 24 objects, drawn on a top-down view of the perceived height map, darker cells higher. Everything shown is computed from the cameras; the object mesh is never read.

From Figure 5: Keyframes of one real trial per object (clamp, blue tape, foldable helix screwdriver), with successes ... — From Figure 5: Keyframes of one real trial per object (clamp, blue tape, foldable helix screwdriver), with successes out of three trials. The screwdriver trial shown failed.

Abstract

Flipping is a fundamental robotic manipulation skill that requires reorienting a target object in a stable pose on a planar support surface by rotating it 180$^\circ$ about an axis parallel to the support surface -- think of flipping a coin from heads to tails. Although flipping is a common objective in industrial tasks such as parts inspection, mechanical assembly, and tool use, coding a bimanual robot with grippers to flip diverse objects is a significant challenge. In this paper, we show that a coding agent can autonomously develop and refine object-flipping strategies through iterative learning in physics simulations across diverse object geometries. Progressive skill acquisition, from basic handover to toppling and pushing, robust grasping and sensing, and thin-object manipulation with recovery, improves task success from 24.7% to 95.0%. The resulting graph-based code integrates high-level strategy selection with reactive control, enabling multi-stage flipping on physical robots.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong