AI

VLAff Learns Where, How, and Which Way From Ego Video

Shar Hendrix 5 min read

Human video is plentiful. Robot action labels are not. VLAff, accepted at IROS 2026 and posted on arXiv August 5, 2026 (arXiv:2608.05215) from the University of Tokyo’s JSK lab, bets that object-centric actionable affordances — where to touch, how to grasp, how to move — can bridge that gap without cloning human joints onto a gripper.

VLAff overview: ego affordances, generation, and real robot tasks
From ego video affordances to robot execution. Source: arXiv:2608.05215 HTML.

EgoAffordance + one model

The pipeline runs SfM, hand-object detection, segmentation, MANO hand reconstruction, ego inpainting, and 3D hand tracking on egocentric footage. The resulting EgoAffordance set holds 204K episodes, about 5.6M visual heatmaps and 11.6M grasp/trajectory affordances (paper abstract; Table I also lists 5.78M heatmaps and 11.61M trajectories).

VLAff extends a Qwen2.5-VL backbone with special tokens: <SEG> for heatmaps (plus DINOv2 features), <GRASP> for 96-D MANO-style parameters, and discrete trajectory bins for autoregressive 6D waypoints. Object-centric transforms pin grasp and path to a 3D anchor from the heatmap peak so outputs stay embodiment-agnostic.

What the numbers say

On visual affordance metrics, VLAff leads the paper’s comparison (IoU 0.121, NSS 1.542, SIM 0.142, KLD 2.517 vs LISA, UAD, VRB, 3DOI).

Zero-shot manipulation: 83.0% average on 10 sim household tasks vs 85.0% for VidBot (strongest sim baseline in the table). Real kitchen tasks on Fetch and PR2: 68.0% average for VLAff vs 52.0% for VidBot across five tasks (10 trials each). Affordance-guided real RL on Open Fridge learns fastest when all three modalities stay in the reward; drop the visual heatmap and learning collapses.

VLAff architecture with VLM, DINOv2, and three affordance decoders
Architecture: VLM plus visual, grasp, and trajectory decoders. Source: arXiv:2608.05215 HTML.

A Human’s Take

I’m excited when “see / grasp / move” comes out of one model trained on kitchen ego video, not another 10k teleop hours. The real-robot gap over VidBot is the money shot; the remaining miss mode — bad first trajectory direction — is exactly the kind of bug you can hunt with better scene priors. Code and weights are listed as coming soon on the project page, so I’m watching that drop.

Sources