full paper: https://t.co/UWSAW8qphw action chunking project: https://t.co/VV0is9u1cV

Takeaway: • Stable dynamics → Action Chunking • Unstable dynamics → Exploratory data via noise Together, they match DAgger-level performance without interaction.

What if dynamics aren’t stable? Action chunking alone fails. The fix: noise-injected expert demos. Add small, structured noise during data collection → expert corrections teach recovery.

Surprising result: chunking helps even without multimodality, diffusion models, or partial observability. It’s not just representation learning . It’s closed-loop stability!

Key insight: if robot dynamics are open-loop stable (common in manipulation), executing longer action chunks naturally damps errors. Result: horizon-free imitation guarantees ,even for deterministic policies.

Why does action chunking work so well in robot imitation learning? This paper spends 99 pages answering that question—showing, via control theory, that chunking stabilizes behavior cloning and prevents exponential error blow-up.

Paper: https://t.co/3FCxbWb7E5 Code:

An automatic pseudo-labeling pipeline curates 1.4M stereo pairs from internet data to supplement synthetic training, boost generalization. The result is a stereo model that runs over 10x faster than FoundationStereo and matches its accuracy.

Knowledge distillation compresses a heavy teacher model into a student backbone. Block-wise neural architecture search finds cost-filtering networks. Structured pruning trims redundant refinement paths without hurting accuracy.

To accelerate stereo matching, it uses a divide-and-conquer workstream that tweaks each major step: feature extraction, cost filtering, and disparity refinement.

The goal: make high-quality stereo depth fast enough for real systems such as robotics, AR/VR, and autonomous machines. Engineered for real-time inference without sacrificing generalization.

NVIDIA just solved the biggest bottleneck in robot vision !! Fast-FoundationStereo runs 10x faster than the original while keeping the same accuracy with real-time 3D depth mapping that actually works anywhere So no more choosing between speed and quality !

Paper:

Key insight: Training on large-scale video gives robots access to motion and causality that static images miss, making Video-Action Models a stronger foundation for generalization

mimic-video is validated on both simulated and real robots, including bimanual systems. It consistently outperforms prior methods in generalization and robustness, especially in low-data regimes where VLAs break down.

Key result: ~10× better sample efficiency and ~2× faster convergence than VLA baselines. Robots learn manipulation skills with far less expert data, lowering the cost of real-world robot training.

The paper introduces mimic-video, a Video-Action Model (VAM). It pairs a large video generative model with an inverse dynamics decoder, letting robots infer actions directly from predicted future motion rather than static semantics.

Vision-Language-Action (VLA) models struggle with physical understanding because they rely on static images. This paper argues robots need temporal priors. Video models naturally encode motion, dynamics, and cause-effect - critical for real control.

Why this matters for robotics: This is the same actuator + controller stack robots use, pushed to hundreds of axes. Deterministic timing. Repeatable trajectories. Reliable motion at scale. This is how better robots get built!

Looks like football ⚽️ It’s actually a motion control demo. IAI’s Magic Field runs on 300+ electric linear actuators under a fixed surface, all synchronized to move a ball with millimeter accuracy. No robots on the field, just very tight control.