0 and RoboDojo, including Video-VAE latents from generative models and raw DINO features.
Control objectives, not reconstruction quality, should guide representation design in world-action models
The authors show that perceptual features like DINO, when calibrated and shaped by action-loss gradients, enable effective joint policy learning and world prediction without generative video pre-training.
Top University
Haoyi Jiang · Liu Liu · Xinjiang Wang · Zhihao Sun · Zequn Chen · Sen Wang · +9 more
Huazhong University of Science and Technology · D-Robotics · Horizon Robotics · Fudan University · Xi'an Jiaotong University
Research Digest··2 min read
Through controlled comparisons, the authors demonstrate that reconstruction fidelity alone is insufficient for choosing representations in world-action models, and that perceptual features benefit from calibration for dynamics modeling.
Why this paper
From Fudan University and 4 others
In one line
ReWAM shapes representations with action objectives, achieving 93.6% success on RoboTwin 2.0 without generative video pre-training.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§