EWAM from Sun Yat-sen and Tsinghua: robots spontaneously learn understand-foresee-act layering
青稞AI · wechat · 2026-10-19
Researchers from Sun Yat-sen University, Tsinghua and 7 other institutions present EWAM, a unified embodied model placing a VL expert (Qwen3-VL-2B), a video-foresight expert (Wan2.2-TI2V-5B) and an action expert in one flow-matching DiT, connected only by asymmetric attention masks with no per-layer supervision.
Key finding: specialization emerges spontaneously
- Action attention splits into three phases: layers 0–9 attend 75.4% to VL (understanding), around layer 13 shift to predicted future frames (imagination), layers 20–29 attend 59.5% to actions themselves (refinement)
- The pattern appears within the first few hundred training steps, holds across all 50 RoboTwin tasks, and is causally load-bearing: masking future frames in mid layers crashes success from 92.8% to 3.55%; injecting another trajectory's "future" drops 5-task success from 25/25 to 4/25
Results
- RoboTwin2.0 clean-to-random: 77.2% average, +26.1 points over π0.5 (Fast-WAM drops to 1.9%)
- LIBERO average 98.8%; real-robot average 72.9% across Franka/Dobot/Unitree G1-D vs π0.5's 51.4%
- Human-video pretraining lifts cross-embodiment transfer from 36.2% to 66.9%; 1,000 headset-collected human demos raise pouring success from 10% to 60% at far lower cost than teleoperation
- Subtask/progress heads lift a long-horizon sort+stack task from 1.5% to 16.0%
Code and four checkpoints are open-sourced. The attention-profiling-then-masking method also offers a way to audit whether other multi-expert models actually use their prediction branches.
More from Embodied
- MIT CSAIL's X-Hinges 3D-Prints Everyday Objects With Built-In Sensing Across Three Motion Axes — MIT News AI · 2026-10-23
- Formic raises $50M Series B to scale physical AI across 100+ US factories — garrytan · 2026-10-23
- fal starts fine-tuning a physical robot to draw, results next week — OdinLovis · 2026-10-23
- PEARL demo: an LLM agent that checks experiment results, revises plans, and re-tests — KevinKaichuang · 2026-10-23
- CU Boulder's ATLAS Institute Opens Tenure-Track Robotics Professorship — ryosuzk · 2026-10-23
- MiTaS multi-resolution tactile imitation learning hits 80% success vs 35.5% vision-only baseline — Jan_R_Peters · 2026-10-23