ActionPiece: Rethinking Action Tokenization for Autoregressive Vision-Language-Action Models
Shijie Lian, Bin Yu, Zhaolong Shen, Xiaopeng Lin, Yichao Du, Zhirui Zhang, Laurence T. Yang, Kai Chen
cs.RO, cs.AI, cs.CL, cs.CV
2026-09-16
ActionPiece adds physical rank losses to an RVQ action tokenizer. Same Qwen3-VL-4B policy: 94.8% LIBERO, 68.8% LIBERO-Plus, +1.1/+4.5 vs strongest matched baselines.
Autoregressive vision-language-action models turn continuous robot commands into discrete tokens. The policy learns those tokens; a decoder turns them back into executable motion. Tokenizers do this with coordinate bins, DCT plus BPE (FAST), or learned vector quantization. Fidelity is usually scored with pointwise reconstruction error such as MSE.
Small per-action error does not mean that the adjustments across demonstrations survive. Similar motions recur with small changes in object pose, robot state, and execution stage. After compression those motions can collapse onto a representative trajectory, and the adjustments needed in a new context get shrunk, warped, or reversed. If the policy predicts the right token sequence, the robot still executes the reconstructed chunk. That is a control problem, not a cosmetic reconstruction problem.
Physical rank consistency (PRC) scores relational fidelity. For each action chunk it takes the k nearest neighbors in the original action space (k=32) and compares physical-distance rankings to those same neighbors before and after reconstruction, using Spearman correlation. Physical distance mixes translation, shortest geodesic rotation on SO(3), and gripper state. Measuring in decoded action space makes token vocabularies and decoder architectures comparable.
ActionPiece is a shallow Transformer codec with residual vector quantization. An 8-step chunk becomes 16 discrete tokens, codebook size 512, quantization depth Q=1. Two rank losses sit on top of MSE and codebook commitment.
Both extra losses are weighted at 6.25×10^{-4}; commitment is 0.25. After tokenizer training the encoder and quantizer freeze, their indices join the VLM vocabulary, and the policy is trained with ordinary next-token prediction.
The matched comparison fixes Qwen3-VL-4B, data, prompt, batch 128, and an 8-step prediction and execution horizon. LIBERO-Plus is never used in training; it tests camera, robot, language, lighting, background, noise, and layout shifts.
| Tokenizer | LIBERO | LIBERO-Plus |
| ActionCodec | 93.7 | 64.2 |
| FAST | 92.1 | 64.3 |
| Standard RVQ | 90.9 | 60.4 |
| ActionPiece | 94.8 | 68.8 |
The out-of-distribution gap is larger: 68.8 versus 64.3 for FAST, with a lead on six of seven shift types. Ablations: PRP alone 93.8/65.4, QR alone 92.9/62.4, both 94.8/68.8. On camera shift, PRP alone is 35.5% and the combination is 45.1%. PRC rises from 0.902 for standard RVQ to 0.953.
On SimplerEnv WidowX the four-task average is 71.9%, above LangForce at 66.5% and ActionCodec-BAR at 65.2%. Block stacking is strongest at 58.3%; putting a carrot on a plate is 54.2%, below π0.5 at 64.7%. On VLA-Arena, trained only at L0, success is 82.2%/42.7%/29.5% on L0/L1/L2, equal-weight 51.5% over 33 cells. Across 55 tokenizer-benchmark pairs, PRC correlates with success at Spearman 0.681 versus 0.544 for reconstruction error.
Swapping an action tokenizer is cheaper than swapping a whole policy. The in-distribution gain is 1.1 points; the unseen-shift gain is 4.5. That pattern fits "stop crushing the adjustments" more than a leaderboard blowout. The matched tokenizer table is the result to trust. SimplerEnv mixes different policy heads and recipes. Anyone shipping a discrete action interface should report PRC next to MSE.
There is no standalone limitations section. SimplerEnv and VLA-Arena compare different policies, so they are not a fair tokenizer-only contest. The carrot task shows physical-order supervision does not win every precise placement. Only Qwen3-VL-4B was used. Quantization depth Q=1 makes the RVQ label mostly branding. There is no real-robot closed loop; SimplerEnv trains on BridgeData and scores in simulation. The extra losses are tiny, and the paper does not test whether those weights stay stable under a different action distribution.