Research map / Reinforcement learning / Visual and representation RL
PRISE: LLM-Style Sequence Compression for Learning Temporal Action Abstractions in Control
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PRISE, which quantizes continuous actions into discrete codes with a state-conditioned VQ module and applies BPE to learn variable-horizon skill tokens that improve behavior cloning on MetaWorld and LIBERO.
Problem
Learning temporal action abstractions (variable-timespan skills) from multitask robotic demonstrations to improve downstream behavior cloning in continuous control with pixel observations.
Approach
Treat temporal action abstraction as sequence compression: discretize continuous actions via a state-conditioned vector-quantized action encoder trained with a latent forward-dynamics (BYOL-style) objective, then apply byte pair encoding over the code sequences to induce variable-length skill tokens used by a high-level BC policy.
Key results
- On LIBERO 5-shot unseen tasks, PRISE achieves 52.7% mean success vs 42.0% for the best baseline (ACT pretrained), a 9.2-point improvement noted in the text.
- On MetaWorld 5-shot unseen tasks, PRISE reaches 66.8% mean success vs 59.3% for ACT (pretrained) and 60.1% for PRISE w/o BPE.
- On LIBERO-90 multitask learning, PRISE is reported to substantially outperform the BC baselines and ACT, with the main architecture matching ResNet-T plus quantization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PRISE skill tokens instead of removing BPE temporal skill tokens (ablation) | PRISE on MetaWorld 5-shot IL · MetaWorld 5-shot IL, 5 unseen tasks | success rate: 60.1± 1.5 -> 66.8± 1.8 | Helped |
| Use PRISE skill tokens instead of removing BPE temporal skill tokens (ablation) | PRISE on LIBERO 5-shot IL · LIBERO 5-shot IL, 8 unseen tasks | success rate: 43.5± 0.6 -> 52.7± 0.9 | Helped |
| Add the latent forward-dynamics objective (ablation) | PRISE action quantization · LIBERO action-code separation | ζ: 14.53 -> 72.42 | Helped |
| Use the PRISE-pretrained encoder for BC | BC on MetaWorld · MetaWorld 5-shot IL, 5 unseen tasks | success rate: 17.3± 4.2 -> 57.9± 1.4 | Helped |
| Use the PRISE-pretrained encoder for ACT | ACT on MetaWorld · MetaWorld 5-shot IL, 5 unseen tasks | success rate: 55.0± 1.8 -> 59.3± 2.0 | Helped |
| Use the PRISE-pretrained encoder for BC | BC on LIBERO · LIBERO 5-shot IL, 8 unseen tasks | success rate: 40.4± 1.2 -> 39.0± 4.1 | Hurt |
| Use the PRISE-pretrained encoder for ACT | ACT on LIBERO · LIBERO 5-shot IL, 8 unseen tasks | success rate: 41.5± 2.1 -> 42.0± 3.2 | No clear effect |
| Pretrain PRISE skills with more trajectories per task (ablation) | PRISE downstream adaptation · LIBERO 5-shot IL | success rate | Helped |
| Vary BPE vocabulary size across the reported range (ablation) | PRISE downstream adaptation · LIBERO 5-shot IL, 8 unseen tasks | success rate | Mixed |
| Increase the number of action codes to 40 or 50 (ablation) | PRISE temporal abstraction · LIBERO 5-shot IL | success rate | Hurt |
Surprising findings
- Pretrained CNN encoder from PRISE did not improve BC or ACT on LIBERO, consistent with reported negative transfer from LIBERO-90 pretraining, while pretrained skill tokens still helped.
- Pretrained encoders from PRISE improved BC and ACT on MetaWorld but not on LIBERO.
Limitations
- Evaluated only on simulated manipulation benchmarks (MetaWorld, LIBERO); real-robot results not shown.
- Performance sensitive to vocabulary size and number of action codes; very large or small vocabularies degrade results.
- Requires pretraining data with meaningful repeated behavioral patterns; gains on LIBERO depend on pretraining data size (drops below 30 trajectories per task).
Experimental setup
Datasets: LIBERO-90 LIBERO-LONG MetaWorld
Benchmarks: LIBERO-90 LIBERO-LONG (8 unseen tasks, 5-shot) MetaWorld (5 unseen tasks, 5-shot)
Metrics: success rate
Baselines: BC (ResNet-RNN) BC (ResNet-T) BC (ViT-T) ACT PRISE w/o BPE BC from scratch
Builds on: ACT VQ-VAE Byte Pair Encoding BYOL ResNet-T
Closest papers by meaning
- QueST: Self-Supervised Skill Abstractions for Learning Continuous Control (NeurIPS 2024 · similarity 0.89)
- Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling (ICLR 2025 · similarity 0.87)
- BEAST: Efficient Tokenization of B-Splines Encoded Action Sequences for Imitation Learning (NeurIPS 2025 · similarity 0.86)
- Think Small, Act Big: Primitive Prompt Learning for Lifelong Robot Manipulation (CVPR 2025 · similarity 0.86)
- Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.86)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.