First promising result of a trained value function for RECAP-style advantage conditioning.
I used a frozen SigLIP-2 to encode the camera views + proprio, with a small MLP value head (d=512) on top.
Aurel Arnold
@aurel_arnold
Building ML models for robotic arms at Dream Machines. Robotics MSc @ETH, prev Serenity @leggedrobotics. Into VLAs, RL.
- I know real-world evals for robot policies are exhausting, but how can you draw conclusions from 12 episodes for a task like folding crumpled shorts? I feel like watching a deployed policy often gives you a good intuition about its performance without needing a huge number of
- PLA-CF is very stiff but brittle. The Dream-Grippers broke frequently when grasping objects with too much force [1]. Printing them in CF-reinforced nylon (PA-CF) makes them much more durable [2]. Some learnings from printing PA-CF: - dry your filament if you don't want a
- How should observations be aligned in robot learning? We record every modality at its native rate: joint states at much higher frequency than images. For training we currently resample proprioception to the image timestamps. At inference we just take the newest of everything.
- 37 / 40 successful rollouts, the jittery motion is fixed! Our implementation of RTC (real-time chunking) had a bug where the number of actions executed during policy inference was estimated from wall-clock time (⌈inference_latency × fps⌉), which was off and led to jumpsOur policy is getting pretty good. 88% success rate on actuator unboxing.
