SJTU's Value-as-Memory Gives Robot Policies Task-Progress Memory, +29.4pp on Real Tasks
jiqizhixin · x · 2026-11-15
Shanghai Jiao Tong University, Shanghai Innovation Institute and Noematrix released an arXiv preprint, Value as Memory (VaM), tackling a core robot-learning problem: policies struggle to tell whether a nearly identical camera view means real progress or a failed grasp.
- Bottleneck: limited visual histories hide whether an action worked, while offline-only training delays feedback from situations a deployed policy actually encounters.
- Mechanism: a value expert estimates future returns, and the action expert reads cached recent value predictions alongside vision and robot state. Training runs demonstration initialization and offline RL warmup, then online human corrections (DAgger-style), then autonomous experience learning — with collection and training running concurrently.
- Results: across box assembly, laundry folding and magnetic bus assembly, value caching adds 29.4 percentage points.
The architecture diagram shows a VLM backbone driven by a history vision encoder, a value cache, and advantage tokens feeding an Action Transformer.
More from Embodied
- ROS 2 Record-and-Replay for the SO-101 Arm Using MCAP and Foxglove — kamathsblog · 2026-11-15
- He reverse-engineered his smart glasses into a live AI object-recognition rig with Opus and YOLO — IgorCarron · 2026-11-15
- Genesis Team Launches Eden, a Robot Learning Framework That Builds Tasks 1.6-4x Faster Than Isaac Lab — zhou_xian_ · 2026-11-15
- Latent-space video world model beats V-JEPA 2, DINO-WM and Dreamer on real gas-station drive-offs — richdotca · 2026-11-15
- Growing Bans on AI Smart Glasses Show Cultural Response Powering Privacy Fight — CarissaVeliz · 2026-11-15
- China flies 10-ton electric passenger plane with pilot aboard after autonomous tests — MarcoFigueroa · 2026-11-15