← Paper Digest

GraSLA: A Lightweight Spatial-Language-Action Model with Grasp Hypotheses as Geometric Priors

Jibao Yuan, Zixiang Zhao, Wensong Xu, Siqi Lu, Shaofei Chen

November 2026

TL;DR GraSLA feeds grasp candidates from a frozen 6-DoF grasp detector, together with point-cloud and language tokens, into a 0.3B-parameter behavior-cloning policy that outputs continuous action chunks. It reaches 99.3% average success on the four LIBERO suites at 54.3 ms per action chunk, in simulation only.

What they do

Policies that plan to a predicted grasp pose can choose among many grasps but execute open-loop. Closed-loop continuous-action policies react to the scene but are limited by how varied their demonstrations are. GraSLA combines the two: grasp hypotheses from a 6-DoF grasp detector serve as geometric priors for a language-conditioned policy trained by behavior cloning.

How it works

  • Three token streams. Utonia encodes multi-view colored point clouds into spatial tokens, and BERT encodes the instruction into language tokens. GSNet proposes grasps from the wrist-view point cloud, and after collision checking each grasp’s pose, gripper configuration and local features become a grasp token.
  • Fusion. Stacked spatial-language-grasp blocks apply bidirectional cross-attention, first between spatial and language tokens and then between grasp tokens and both, followed by a language enhancement module.
  • Actions. An ACT-style decoder reads the fused tokens and the robot state and predicts an action chunk, trained with an L1 loss. BERT and GSNet stay frozen.

Results

  • LIBERO (Spatial, Object, Goal and Long suites, 50 rollouts per task): 99.3% average success. The best other 3D policies score 99.2% (Flex-π*) and 99.0% (PointmapVLA), and both pretrain on additional robot demonstrations.
  • Cost: 0.3B parameters and 54.3 ms latency on an RTX 5090, versus 6.0B and 590.7 ms for Flex-π*.
  • Ablations: removing the grasp branch lowers the average to 98.1% (LIBERO-Long 98.8% to 97.0%). Removing the spatial or the language branch drops it to 84.15% and 68.75%.

Limitations

Evaluated only in simulation on LIBERO, where several baselines already exceed 98%. Behavioral diversity, the paper’s motivation, is not measured; only Figure 1 illustrates it. Grasp hypotheses are not conditioned on the instruction and describe only parallel-jaw grasps, so they give little guidance for pushing or pressing.

Figures

From Figure 1: Comparison of manipulation pipelines: (a) pose-based methods, (b) continuous-control methods, and (c) ...
From Figure 1: Comparison of manipulation pipelines: (a) pose-based methods, (b) continuous-control methods, and (c) GraSLA. (d) Comparison of their behavior.
From Figure 2: Overview of GraSLA. (a) Multi-view point clouds, the language instruction and grasp hypotheses become ...
From Figure 2: Overview of GraSLA. (a) Multi-view point clouds, the language instruction and grasp hypotheses become spatial, language and grasp tokens; the fused tokens and the robot state go to an action decoder that predicts an action chunk. (b) In each interaction block, bidirectional attention updates the spatial and language tokens, then all three streams, and self-attention refines the language tokens.
Abstract

This paper focuses on improving the behavioral diversity and effectiveness of language-conditioned manipulation policies trained through imitation learning. Recently, continuous-action policies have become mainstream in robotic manipulation, but the lack of diverse demonstrations tends to restrict their behavioral diversity. For grasping, a fundamental subtask of manipulation, 6-DoF grasp detection provides diverse grasp hypotheses, but open-loop execution constrains responsiveness in dynamic environments. This motivates us to combine the diversity of grasp hypotheses with the adaptability of continuous-action policies. To this end, we propose GraSLA, a lightweight and effective spatial-language-action model with grasp hypotheses as geometric priors. In particular, we employ spatial and language foundation models to encode scene geometry and task semantics, and integrate them with grasp hypotheses through bidirectional cross-attention for continuous action prediction. On LIBERO, GraSLA achieves an average success rate of 99.30%.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.