Driving · 2026-11-10
SPARK: Simple Post-training for Adapting pRetrained Knowledge to Robot Control
Taekyung Kim, Jeongeun Park, Byeongho Heo, Sungjoon Choi, Sangdoo Yun, Dongyoon Han
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
SPARK is a post-training method that adapts pretrained vision/foundation models for robot control. It uses temporal abstraction to encourage compact features for temporally offset scene prediction and knowledge preservation to align patch-level representations with the original foundation model, aiming to improve success rates and generalization on robot manipulation tasks and transfer to real-world robot manipulation.
Why it matters
The authors address a mismatch between foundation model objectives (global semantics, not temporally sensitive representations) and robot control needs, and they avoid training robot representations from scratch by post-training pretrained models. SPARK is designed to improve control performance while retaining pretrained semantic knowledge, and is evaluated on multiple simulated benchmarks and real-world robot manipulation tasks.
Method
- Temporal abstraction: train an encoder to produce a compact class token from a reference observation that supports prediction of a temporally offset target scene via token reconstruction with visible target hints.
- Knowledge preservation: regularize adaptation by aligning adapted patch-wise representations with those from a frozen original foundation model on the target scene.
- Post-training objective combines temporal abstraction and knowledge preservation losses to obtain visual state representations suited for control without training from scratch.
Abstract (from arXiv)
Large-scale representation learning has produced public foundation models that provide strong general-purpose visual features. However, their pretraining objectives are not designed for robot representation learning, so the resulting embeddings tend to emphasize global semantics rather than the temporally sensitive representations required for robot control. Training robot models from scratch is also undesirable due to the limited scale and high collection cost of robotics data, as well as the substantial computational burden of pre-training. In this paper, we propose SPARK (Simple Post-training for Adapting pRetrained Knowledge), a post-training method that adapts foundation models to robot control. SPARK is built on two principles: temporal abstraction, which encourages the encoder to derive compact features that preserve information essential for understanding temporal changes and action-relevant scene structure, and knowledge preservation, which aligns patch-level representations with those of the original foundation model to retain useful pretrained semantics. These objectives yield compact visual state representations that remain semantically meaningful while preserving useful prior knowledge. Experiments across multiple robotics benchmarks show that SPARK consistently improves success rates and generalization over vanilla foundation models, and further demonstrate that these gains transfer to real-world robot manipulation.
Related papers
- SLA: Scaling Spatial-Language-Action Model via Continual Pre-training
- Plan First, Act Later: Streaming Reasoning for Executable Memory in Robot Control
- LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies
- RoboAtom: Large-scale Atomic-Action Supervision for Robot Learning
- Video2STL: Grounding VLM-Generated Temporal Specifications for Robot Learning
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-11-10 · Indexed by Pub-AI: 2026-11-10 UTC
SPARK: Simple Post-training for Adapting pRetrained Knowledge to Robot Control
First announced by arXiv: 2026-11-10 · Original record ↗
- arxiv
- 2611.12102
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-11-10 · First detected: 2026-11-10 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.12102v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
Large-scale representation learning has produced public foundation models that provide strong general-purpose visual features. However, their pretraining objectives are not designed for robot representation learning, so the resulting embeddings tend to emphasize global semantics rather than the temporally sensitive representations required for robot control. Training robot models from scratch is also undesirable due to the limited scale and high collection cost of robotics data, as well as the substantial computational burden of pre-training. In this paper, we propose SPARK (Simple Post-training for Adapting pRetrained Knowledge), a post-training method that adapts foundation models to robot control. SPARK is built on two principles: temporal abstraction, which encourages the encoder to derive compact features that preserve information essential for understanding temporal changes and action-relevant scene structure, and knowledge preservation, which aligns patch-level representations with those of the original foundation model to retain useful pretrained semantics. These objectives yield compact visual state representations that remain semantically meaningful while preserving useful prior knowledge. Experiments across multiple robotics benchmarks show that SPARK consistently improves success rates and generalization over vanilla foundation models, and further demonstrate that these gains transfer to real-world robot manipulation.
