Tsinghua team's ActionCodec2 streams action tokens per step, hitting 46.2% grasp vs FAST's 9.2%
ZeYanjie · x · 2026-10-31
Researchers from Tsinghua University and Galaxea AI released ActionCodec2, a streaming action tokenizer that fixes the chunked execution pattern of today's VLAs — plan a second, then run blind.
Key ideas
- Emits tokens for every control step instead of encoding 1–2s chunks, so the robot acts as each token arrives and can react mid-execution
- Frames tokenization as a covering problem: a fixed grid covers each step's reachable motion, guaranteeing a bounded error (1/5/10 mm options) even on unseen motions
- Set-BPE losslessly merges co-occurring tokens to keep streams short
- Fitted on 24 embodiments with 1k–16k vocabularies, drop-in ready for any VLA, no extra neural network needed
Results
- Real-robot grasping: closed-loop ACodec2+ reaches 46.2% grasp rate vs 27.6% open-loop and 9.2% for FAST
- Disturbance tests: when objects move mid-chunk, the closed-loop version re-reads state each step and still grasps, while open-loop methods fail on stale chunks
Paper, GitHub and Hugging Face weights are all open, with a three-line quickstart on the project page.
More from Embodied
- AgiBot Sets Guinness Record With 110 Humanoid Robots Dancing in Sync — emmanuelvivier · 2026-10-31
- David Lieb's reaction to subvocal wristband demo goes viral; silent input seen as step toward telepathy — LinusEkenstam · 2026-10-31
- Engineer shows 10+ years of robot dev board evolution, from bulky to compact — k7agar · 2026-10-31
- RL-trained robot arm keeps a ping-pong ball bouncing on a paddle for 7 minutes — ___Mufasaa · 2026-10-31
- Two 256GB Mac Studios beat one 512GB for local LLMs: 2.2TB/s effective bandwidth — lxfater · 2026-10-31
- Robotics sim buzz: Sentdex and Stone Tao praise a new source of simulation assets — AIFlow_ML · 2026-10-31