AI

GIFT Trains Robot Policies to Keep Geometry, Affordances, and Goals

Robb Harlan 4 min read

A team from the Chinese Academy of Sciences, Tsinghua, Fudan, and the National University of Singapore posted GIFT on 3 September (arXiv:2609.04193): a training recipe that forces a robot policy’s intermediate visual features to keep three control-relevant structures. Geometry for “can I move there,” affordance for “what to grab and how,” and goals for “which pixels actually matter for the instruction.”

The authors call the usual mismatch the action-sufficiency gap. Vision-language pretraining and world models give rich pictures. Those pictures still dump a lot of stuff a gripper does not need.

UFACTORY xArm 7 fruit-shelf setup and dual ARX X5 lab bench with cameras labeled
Real evaluation rigs: single-arm UFACTORY xArm 7 and dual-arm ARX X5. Source: Zheng et al. / GIFT project page.

What they actually train

GIFT does not invent a new action head. It wraps three existing policy families and keeps each family’s action math:

  • GIFT-VLA: a semantics-centered vision-language-action policy
  • GIFT-WAM-Fast: a world-action model that dumps a current-frame action
  • GIFT-WAM-IDM: an inverse-dynamics world-action model that imagines a trajectory first

The extra losses sit on intermediate tokens. Geometry is aligned to a frozen VGGT teacher. Affordance heads predict entity roles, object-centric end-effector poses, and gripper closure. A goal head reconstructs instruction-relevant image regions. At deployment the extra heads can stay off. The paper’s default is no-injection: the auxiliaries supervise features, they do not get concatenated into the action path.

On zero-shot LIBERO-Plus (seven distribution shifts, no fine-tune), the project page reports:

  • GIFT-VLA 79.6% vs StarVLA-OFT 75.0% (+4.6 points)
  • GIFT-WAM-Fast 72.6% vs Fast-WAM 60.0% (+12.6)
  • GIFT-WAM-IDM 87.8% vs Fast-WAM-IDM 82.6% (+5.2)

On RoboCasa, the same three land at 61.4%, 83.6%, and 82.3%. Articulated-object tasks are where the world-action variants move the most: +21.3 and +24.6 points over their Fast-WAM counterparts.

Grid of dual-arm rollouts placing a cube in a cabinet and inserting test tubes, with and without visual perturbations
Dual-arm rollouts, including unseen lighting and tabletop clutter. Source: GIFT project page.

Hardware, not just sim

The real suite is four tasks, ten trials each, on the two platforms above: color-conditioned placement, layer-conditioned placement, cabinet work, and bimanual test-tube insertion. GIFT-WAM-IDM hits 87.5% in the original settings and 67.5% across perturbation conditions, against 52.5% and 15.0% for Fast-WAM-IDM (40 original trials, 40 perturbed). Level-2 perturbations include a rotating colored lamp, a swapped tabletop, and a 15–20° rack rotation.

Ablations on the project page say each single guidance signal helps, all three together win, and the no-injection variants beat the injection ones on both LIBERO-Plus and RoboCasa.

Single-arm color- and layer-conditioned placement sequences under rotating lamp and changed tabletop
Single-arm placement under lighting and background shifts that were not in training. Source: GIFT project page.

A Human’s Take

I like that they left the action decoder alone. If your VLA already works, the useful question is whether the features it sees are the ones that decide a grasp, not whether you can stack another transformer on top. The real-arm gap under perturbation is the number I would take into a lab meeting. Eighty-seven percent on a clean table is nice. Sixty-seven percent when someone spins a lamp and rotates the rack is closer to a shift.

Sources