GRITS: A Spillage-Aware Guided Diffusion Policy for Robot Food Scooping Tasks

Authors: Yen-Ling Tai, Yi-Ru Yang, Kuan-Ting Yu, Yu-Wei Chao, Yi-Ting Chen · ICRA 2026

Licences

Open Access
Paper: CC BY 4.0

Everything below draws on the full paper under its CC BY 4.0 licence; its figure is credited to the authors.

TL;DR

GRITS guides a diffusion policy at test time with a spillage classifier trained on 4,000 simulated scoops, cutting real-world spillage from 45% to 4% on unseen foods.

Why it matters

For food-preparation and assistive-feeding robots, spillage is a safety and cleanliness failure, not only a lower score. GRITS adds this objective at test time without retraining the policy, and learns it from simulation instead of from real spills.

Problem

Learning-from-demonstration scooping policies only cover the food states seen in their demonstrations, and diffusion policies that imitate well still spill when they meet unseen states such as a very full bowl. Teaching a policy to avoid spillage with real examples is impractical, because staging spills takes heavy setup and cleanup and risks damaging food.

Proposed solution

A CNN-based diffusion policy is trained on 80 kinesthetic demonstrations, and a separate spillage predictor is trained on 4,000 simulated scooping trials in Isaac Lab built from spheres, cubes, cones and cylinders with randomised mass, friction, radius and quantity. At inference the predictor scores the estimated clean trajectory against a segmented point cloud (SAM2-segmented food plus known spoon and bowl models, encoded with DP3), and its gradient, weighted by ρ = 2.5, is subtracted during denoising after a threshold step s = 30 so a coarse trajectory forms first.

Key insights

A differentiable classifier trained only in simulation can guide a real-robot diffusion policy when its input is a segmented point cloud rather than RGB.
Applying the spillage objective inside the denoising loop works better than correcting trajectories afterwards: post-hoc correction lowered spillage but raised scoop failure to 40%.
Composite point clouds (camera-reconstructed food plus known spoon and bowl models) were critical: raw sensor point clouds gave 61% success and 23% spillage.
Guidance trades a little scoop failure for safety: 14% scoop failures with GRITS versus 10% for the unguided diffusion policy.
Delaying guidance until a coarse trajectory exists (threshold s) avoids over-correction in early denoising steps.

Main results

Ten unseen food categories, two quantity levels, five trials each: GRITS reached 82% task success, 4% spillage and 14% scoop failure.
Unguided diffusion policy: 48% success and 45% spillage; SCONE: 52% and 17%; behaviour cloning: 46% and 43%; rule-based: 29% and 50% (Fig. 6).
Diffusion policy with post-hoc spillage correction: 45% success and 40% scoop failure; its spillage rate is 8% in the text but 17% in Fig. 6.
Spillage-predictor ablation: raw point cloud 61% success and 23% spillage; PointNet++ encoder 70% and 17%; composite point cloud with DP3 encoder 82% and 4%.
Per food (Fig. 7), GRITS removed most spillage on sago and red bean; mentos still spilled occasionally, and marshmallow, fudge and prune stayed difficult.
Refinements are small: an average displacement of 0.3 cm between consecutive trajectory points (Fig. 1).

Limitations

Failures often come from food properties that vision cannot capture, such as hardness, viscosity and deformability; the authors propose pre-interaction strategies and force-torque or tactile sensing.
Guidance occasionally misses food collection, raising scoop failure slightly above the unguided policy (14% versus 10%).
Only the scooping stage is learned; transfer uses bowl detection with motion planning, and pouring uses predefined trajectories.
Evaluation is 100 real trials on one Franka setup, with the source bowl fixed and the target bowl placed within a 35 × 30 cm workspace.
Soft, sticky or densely packed foods (marshmallow, fudge, prune, mentos) still caused failures or spillage.

Field impact

Finalist for the ICRA 2026 Best Paper Award on Robot Learning, with 3 citations. The paper does not state a release of code, data or models.

Abstract

Robotic food scooping is a critical manipulation skill for food preparation and service robots. However, existing robot learning algorithms, especially learn-from-demonstration methods, still struggle to handle diverse and dynamic food states, which often results in spillage and reduced reliability. In this work, we introduce GRITS: A Spillage-Aware Guided Diffusion Policy for RobotIc Food Scoop TaskS. This framework leverages guided diffusion policy to minimize food spillage during scooping and to ensure reliable transfer of food items from the initial to the target location. Specifically, we design a spillage predictor that estimates the probability of spillage given current observation and action rollout. The predictor is trained on a simulated dataset with food spillage scenarios, constructed from four primitive shapes (spheres, cubes, cones, and cylinders) with varied physical properties such as mass, friction, and particle size. At inference time, the predictor serves as a differentiable guidance signal, steering the diffusion sampling process toward safer trajectories while preserving task success. We validate GRITS on a real-world robotic food scooping platform. GRITS is trained on six food categories and evaluated on ten unseen categories with different shapes and quantities. GRITS achieves an 82% task success rate and a 4% spillage rate, reducing spillage by over 40% compared to baselines without guidance, thereby demonstrating its effectiveness.

Open Source Robotics