ThinkProprio: Embodied Visual Reasoning for VLA Manipulation

Highlights

Proprioceptive Token Selection

Where the robot is tells us what to look at. Joint instruction-proprioception guidance selects task-relevant visual patches.

58% Faster Inference

Latency drops from 52ms to 22ms by retaining only 15% of visual tokens while maintaining or improving performance.

CALVIN State-of-the-Art

4.55 average completion length on ABC→D with 82.1% success on 5-task chains (LH-5).

Text-Tokenized Proprioception

Discretizing proprioception into VLM tokens outperforms MLP projection, leveraging the pretrained token embedding space.

Method

ThinkProprio introduces proprioception as a first-class modality in VLA pipelines. We text-tokenize the robot's proprioceptive state (joint angles, end-effector pose) by discretizing continuous values into bins and mapping them to VLM vocabulary tokens.

The tokenized proprioception, combined with the task instruction, guides physically grounded token selection: a cross-attention mechanism scores visual patches based on their relevance to both the instruction and the robot's current configuration. A vote-based selection retains only the most relevant patches, with a global context token preserving coarse scene information.

The compact token set is processed by the VLM, and a flow-matching action head generates continuous action chunks via cross-attention over the fused features.

Results

CALVIN ABC→D

Method	LH-1	LH-2	LH-3	LH-4	LH-5	Avg. Len.
OpenVLA	91.3	77.8	62.0	52.1	43.5	3.27
GR-1	85.4	71.2	59.6	49.7	40.1	3.06
π₀	70.0	48.0	37.0	28.0	18.0	2.01
π_0.5	71.0	56.0	45.0	37.0	29.0	2.38
VPP	95.7	91.2	86.3	81.0	75.0	4.29
Seer	96.3	91.6	86.1	80.3	74.0	4.29
FLOWER	99.3	96.0	90.3	82.3	75.5	4.44
ThinkProprio (Ours)	97.7	96.1	92.2	86.7	82.1	4.55

CALVIN ABCD→D

Method	LH-1	LH-2	LH-3	LH-4	LH-5	Avg. Len.
Diff-P-CNN	86.3	72.7	60.1	51.2	41.7	3.16
RoboFlamingo	96.4	89.6	82.4	74.0	66.0	4.09
GR-1	94.9	89.6	84.4	78.9	73.1	4.21
FLOWER	98.9	96.7	93.9	90.2	85.5	4.62
FLOWER†	99.2	96.9	96.9	92.3	88.3	4.67
ThinkProprio (Ours)	99.5	97.2	96.6	92.3	88.5	4.74

CALVIN D→D

Method	LH-1	LH-2	LH-3	LH-4	LH-5	Avg. Len.
MDT	93.7	84.5	74.1	64.4	55.6	3.72
RoboUniView	96.2	88.8	77.6	66.6	56.3	3.85
FLOWER†	97.4	92.4	86.9	81.3	74.9	4.35
ThinkProprio (Ours)	96.9	89.8	83.6	80.5	72.7	4.23

LIBERO Benchmark Suites

Method	Spatial	Object	Goal	Long	Avg.
OpenVLA	84.7	88.4	79.2	53.7	76.5
OpenVLA-OFT	97.6	98.4	97.9	94.5	97.1
π₀	96.8	98.8	95.8	85.2	94.2
FLOWER	97.5	99.1	96.1	94.9	96.9
LightVLA	98.4	98.4	98.2	94.6	97.4
ThinkProprio (Ours)	97.6	98.4	98.0	95.2	97.3

Inference Efficiency (CALVIN ABC→D)

Method	Visual Tokens	Latency (ms)	VRAM (MB)	Avg. Len.
OpenVLA	256	164	14574	3.27
π₀	256	104	6692	2.01
FLOWER	100	52	1848	4.44
ThinkProprio (Ours)	15	22	1899	4.55

Video Demonstrations

Long-horizon task execution (5 sequential subtasks per trajectory) with token selection visualization.

Trajectory 1: Object Manipulation

0. Grasp and lift the red block

1. Stack the grasped block

2. Pull the handle to open the drawer

3. Remove the stacked block

4. Push the pink block right

Trajectory 2: Environmental Interaction

0. Push the sliding door to the left

1. Pull the handle to open the drawer

2. Use the switch to turn on the light

3. Rotate the red block to the right

4. Grasp and lift the pink block

Token Selection Visualization

Token retention heatmaps across timesteps

Token retention across four timesteps for two tasks. Heatmaps visualize which visual patches are selected based on joint instruction-proprioception guidance. The selection shifts between object-centric and proprioception-centric focus as the task progresses.

Citation

@article{wang2025thinkproprio,
  title={Think Proprioceptively: Embodied Visual Reasoning for VLA Manipulation},
  author={Wang, Fangyuan and Zhou, Peng and Qi, Jiaming and Lyu, Shipeng and Navarro-Alarcon, David and Guo, Guodong},
  journal={arXiv preprint},
  year={2025}
}