What do these badges mean?
- 🚀ShippingCode exists. Multiple GitHub repos already reference this paper — people are building on it.
- 📈ClimbingCitation velocity is rising. Researchers are starting to pick it up.
- 💤QuietPublished but no notable signal yet. Most papers live here — could become anything later.
- 🎭HypeHeavy social buzz but no shipping signal. The counter-signal — defer until Twitter/X data is wired up.
- 🚀Shipping2605.16241·May 15, 2026·~13 mincs.CVcs.AI
Offline Semantic Guidance for Efficient Vision-Language-Action Policy Distillation
Jin Shi, Brady Zhang, Yishun Lu
⭐ 220 stars / 10 repos📚 0 citesELI5A method to shrink large robot control models by having a smaller model learn from a bigger one, using high-level task descriptions (like 'grasping phase' or 'move left') alongside low-level action imitation—then throwing away those descriptions at test time so the small model runs fast alone.
Problem solvedLarge vision-language-action models for robots are too slow and memory-intensive for real-time robot control. This shrinks them 44× smaller while keeping performance, enabling 3× faster inference on actual hardware.
- 🚀Shipping2605.16154·May 15, 2026·~15 mincs.LGcs.RO
Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking
Vaidehi Bagaria, Nikshep Grampurohit, Pulkit Verma
⭐ 166 stars / 10 repos📚 0 citesELI5When training robot policies with RL, most time is spent on gradient calculations for boring parts of the trajectory that the robot already knows. This paper identifies which parts of a trajectory actually matter for learning (where successful and failed attempts differ) and only updates the model on those chunks, making training 2.4x faster.
Problem solvedRobot policy training is slow because gradient computation dominates the cost, yet most gradients flow through trajectory segments the policy has already mastered. You waste compute updating on parts that don't contribute to learning, slowing down RL-based improvement of vision-language-action models.