InterACT

UC Davis; UC Berkeley · 2024-09†

VAVANot disclosedregression

An ACT-family policy that models inter-arm dependencies with hierarchical attention and per-arm decoder synchronization.

Paper

Backbone LLM: None
Vision encoder: ResNet18
Action head: Multi-arm decoder (synchronization blocks) (regression)
Chunk: 50 steps
Control rate: 50 Hz
Architecture (original summary): ResNet18 + hierarchical attention encoder (3 layers per segment·3 layers across segments, per-segment CLS token) + multi-arm decoder (2 parallel paths of 4 layers, 1 synchronization block layer); hidden 512, 8 heads, FF 3200; chunk 50, joint positions·RGB 50 Hz

Pretraining accelerators: N/A
Pretraining duration: N/A
Pretraining steps & batch: N/A
GPU/TPU-hours: Not disclosed
Fine-tuning: Per task: GPU·time·epochs Not disclosed; batch 8, lr 1e-5
Data type: Simulation·synthetic; real-robot teleoperation
Data scale: 50 demos per task
Main datasets: In-house collection (sim and real)
Tasks & evaluation: Sim 3 tasks (Transfer Cube, Peg Insertion, Slot Insertion); real 6 tasks (Slide Ziploc, Thread Velcro, Insert Plug, Click Pen, Sweep, Unscrew Cap)
Robot: Bimanual (model not stated in note)

Reinforcement learning: Not used · Imitation learning (behavior cloning) only
Base control: Fixed · No base control (bimanual)

Key features: An ACT-family policy that models inter-arm dependencies with hierarchical attention and per-arm decoder synchronization. Trained with 50 demos per task.
Notes: Task split correction applied (3 sim + 6 real). Announcement month v1 unverifiable (†, consistent with arXiv ID 2409). Low adoption, included as an ACT-family data point.
Verification: Corrections applied