AI

Arm-wise Compositional Generalization in Dual-Arm Vision-Language-Action Models

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 05:18:24 UTC.

Briefings

This study introduces ACG-Bench, a benchmark for evaluating Arm-wise Compositional Generalization in dual-arm Vision-Language-Action models. Using the shared π_{0.5} backbone, the research compares data augmentation and architectural strategies, identifying AE-VLA as the most effective method. On physical SO101 robots, AE-VLA achieved 39.00% mean success across five unseen conditions, significantly outperforming the strongest baseline at 10.00%. The benchmark includes 23 task-condition pairs across 8 families, with 6 in-domain and 17 unseen compositions testing reordering, synchronization, and cross-task skills.

Key points

• AE-VLA achieved 39.00% mean success on physical SO101 robots across five unseen conditions, compared to 2.94% for Single π_{0.5} and 5.53% for two independently controlled π_{0.5} policies.
• ACG-Bench contains 23 task-condition pairs across 8 task families, with 17 unseen compositions covering reordering, synchronization, and cross-task composition.

Context

Existing evaluations offer limited insight into which training and architectural choices support compositional generalization in dual-arm policies. This benchmark provides a common testbed for studying skill recomposition under different coordination requirements.

Limitations

The study relies on simulation and physical robot data; no additional background on the specific limitations of the evaluation protocol or the physical environment is supplied.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 18/10/2026, 02:14:24 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiccsbjc5oek7zsslpubtg4bo3ol4bjimsphfywap7mlvaccdimhdu