mlsanityData to 14 Oct 2026

Action Tokenizer Matters in In-Context Imitation Learning

arXiv 2503.01206 ↗First posted 3 Mar 2025cs.ROcited by 14

On the map: Offline RL and imitation learning from data → Robot manipulation imitation learning

Abstract (arXiv)

In-context imitation learning (ICIL) is a new paradigm that enables robots to generalize from demonstrations to unseen tasks without retraining. A well-structured action representation is the key to capturing demonstration information effectively, yet action tokenizer (the process of discretizing and encoding actions) remains largely unexplored in ICIL. In this work, we first systematically evaluate existing action tokenizer methods in ICIL and reveal a critical limitation: while they effectively encode action trajectories, they fail to preserve temporal smoothness, which is crucial for stable robotic execution. To address this, we propose LipVQ-VAE, a variational autoencoder that enforces the Lipschitz condition in the latent action space via weight normalization. By propagating smoothness constraints from raw action inputs to a quantized latent codebook, LipVQ-VAE generates more stable and smoother actions. When integrating into ICIL, LipVQ-VAE improves performance by more than 5.3% in high-fidelity simulators, with real-world experiments confirming its ability to produce smoother, more reliable trajectories. Code and checkpoints are available at https://action-tokenizer-matters.github.io/

Key facts. Cited by 14 papers in the map (data through 14 Oct 2026). Placed in Robot manipulation imitation learning, one of 13 directions in Offline RL and imitation learning from data. Builds on 47 of its 53 references that are in the map. First posted on arXiv 3 Mar 2025.

Builds on

47 of 53 references.

Closest ancestors · 9

Same direction, newest first

  1. 2025SKIL: Semantic Keypoint Imitation Learning for Generalizable Data-efficient Manipulation24 Jan
  2. 2024Instant Policy: In-Context Imitation Learning via Graph Diffusion19 Nov
  3. 2024In-Context Imitation Learning via Next-Token Prediction28 Aug
  4. 2024Keypoint Action Tokens Enable In-Context Imitation Learning in Robotics28 Mar
  5. 2024PRISE: LLM-Style Sequence Compression for Learning Temporal Action Abstractions in Control16 Feb
  6. 2023MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations26 Oct
  7. 2023Few-Shot In-Context Imitation Learning via Implicit Graph Alignment18 Oct
  8. 2023Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware23 Apr
  9. 2022Behavior Transformers: Cloning k modes with one stone22 Jun

Other references · 38

  1. 2017Attention Is All You Needcited by 60,361
  2. 2015Deep Residual Learning for Image Recognitioncited by 52,096
  3. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  4. 2020Language Models are Few-Shot Learnerscited by 32,419
  5. 2021Emerging Properties in Self-Supervised Vision Transformerscited by 6,968
  6. 2017Neural Discrete Representation Learningcited by 4,712
  7. 2020Taming Transformers for High-Resolution Image Synthesiscited by 3,253
  8. 2024OpenVLA: An Open-Source Vision-Language-Action Modelcited by 2,887
  9. 2022RT-1: Robotics Transformer for Real-World Control at Scalecited by 2,061
  10. 2018Spectral Normalization for Generative Adversarial Networkscited by 1,636
  11. 2024Octo: An Open-Source Generalist Robot Policycited by 1,484
  12. 2023Open X-Embodiment: Robotic Learning Datasets and RT-X Modelscited by 862
26 more
  1. 2024DROID: A Large-Scale In-The-Wild Robot Manipulation Datasetcited by 852
  2. 2021What Matters in Learning from Offline Human Demonstrations for Robot Manipulationcited by 725
  3. 2025FAST: Efficient Action Tokenization for Vision-Language-Action Modelscited by 665
  4. 2023Language Model Beats Diffusion -- Tokenizer is Key to Visual Generationcited by 640
  5. 2024A Survey on Vision-Language-Action Models for Embodied AIcited by 279
  6. 2024ManiSkill3: GPU Parallelized Robotics Simulation and Rendering for Generalizable Embodied AIcited by 229
  7. 2023Large Language Models as General Pattern Machinescited by 197
  8. 2023RoboAgent: Generalization and Efficiency in Robot Manipulation via Semantic Augmentations and Action Chunkingcited by 169
  9. 2023Large Language Models Are Latent Variable Models: Explaining and Finding Good Demonstrations for In-Context Learningcited by 149
  10. 2024Flow as the Cross-Domain Manipulation Interfacecited by 132
  11. 2022Generative Flow Networks for Discrete Probabilistic Modelingcited by 114
  12. 2022Learning Smooth Neural Functions via Lipschitz Regularizationcited by 86
  13. 2019Learning Action Representations for Reinforcement Learningcited by 75
  14. 2023For SALE: State-Action Representation Learning for Deep Reinforcement Learningcited by 75
  15. 2023Language Models as Zero-Shot Trajectory Generatorscited by 60
  16. 2020Regularizing Action Policies for Smooth Control with Reinforcement Learningcited by 59
  17. 2021LASER: Learning a Latent Action Space for Efficient Reinforcement Learningcited by 32
  18. 2021DECORE: Deep Compression with Reinforcement Learningcited by 29
  19. 2024Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasetscited by 23
  20. 2023Kinematic-aware Prompting for Generalizable Articulated Object Manipulation with LLMscited by 19
  21. 2022Inferring Smooth Control: Monte Carlo Posterior Policy Iteration with Gaussian Processescited by 10
  22. 2024Dynamics as Prompts: In-Context Learning for Sim-to-Real System Identificationscited by 10
  23. 2018Action Representations in Robotics: A Taxonomy and Systematic Classificationcited by 9
  24. 2024VQ-ACE: Efficient Policy Search for Dexterous Robotic Manipulation via Action Chunking Embeddingcited by 7
  25. 2023BeBOP -- Combining Reactive Planning and Bayesian Optimization to Solve Robotic Manipulation Taskscited by 6
  26. 2024EMOTION: Expressive Motion Sequence Generation for Humanoid Robots with In-Context Learningcited by 6

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 3; 2026: 11.

14 papers.

Sort:

By area:

  1. 2025Learning Long-Context Diffusion Policies via Past-Token Predictioncited by 30
  2. 2025Physical Autoregressive Model for Robotic Manipulation without Action Pretrainingcited by 12
  3. 2026Wall-OSS-0.5 Technical Reportcited by 12
  4. 2026UniT: Toward a Unified Physical Language for Human-to-Humanoid Policy Learning and World Modelingcited by 10
  5. 2026VPWEM: Non-Markovian Visuomotor Policy with Working and Episodic Memorycited by 6
  6. 2025RoboSSM: Scalable In-context Imitation Learning via State-Space Modelscited by 5
  7. 2026Neural Implicit Action Fields: From Discrete Waypoints to Continuous Functions for Vision-Language-Action Modelscited by 3
  8. 2026In-Context World Modeling for Robotic Controlcited by 3
  9. 2026Ordered Action Tokens for Visuomotor Policy Learningcited by 3
  10. 2026A Hierarchical Spatiotemporal Action Tokenizer for In-Context Imitation Learning in Roboticscited by 1
  11. 2026ContextFlow: In-Context Flow Matching for Robot Manipulationcited by 1
  12. 2026In-Context Learning for Robots: Methods and Applicationscited by 1
2 more
  1. 2026Unsupervised Skeleton-Based Action Segmentation via Hierarchical Spatiotemporal Vector Quantizationcited by 0
  2. 2026Enhancing Human-Likeness in Reinforcement Learning Agents via Hierarchical Macro Action Quantizationcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025BEAST: Efficient Tokenization of B-Splines Encoded Action Sequences for Imitation Learning
  2. 2026Beyond Reconstruction: What Matters in Action Tokenization for Robot Policies?
  3. 2026WorldToken: Time-First Sequence Modeling for Robotic Imitation Learning
  4. 2026ICI-VLA: In-Context Imitation with Spatiotemporally Aligned Demonstrations for Vision-Language-Action Models
  5. 2026Primary-Fine Decoupling for Action Generation in Robotic Imitation
  6. 2025Latent Action Pretraining Through World Modeling
  7. 2026OAT: Ordered Action Tokenization
  8. 2026Looking Back to Move Forward: Temporal Verification for Generative Robot Policies