01
Procedural reasoning
Explicit task graphs and multimodal memory maintain the active step, completed actions, dependencies, branch conditions, and task-relevant evidence across long procedures.
Task graphsMemoryLong horizon
Research project · Robot learning · Vision-language-action
Reliable visuomotor learning for long-horizon manipulation.
We study how robots can act reliably when success depends not only on learning a manipulation skill, but also on understanding what should happen next and what matters visually right now.
PROCEDURE
What should happen next?
Task graphs · memory · state
GROUNDING
What matters now?
Objects · destinations · phase
VLA
policy
policy
ACTION
Reliable long-horizon control
Execute · verify · continue
Research direction
Two complementary failure modes. One reliability problem.
Long-horizon robot behavior breaks when the system loses procedural state or visually grounds the wrong target. Our work addresses both layers of the problem.
02
Conditional visual grounding
The relevant object or destination changes with the manipulation phase and observed task state. We diagnose distractor failures and improve target selection without discarding spatial control information.
DistractorsAttentionVisual prompting
Publications
ECCV 2026 workshop papers
ECCV 2026 · X-Reason
Paper 01
Oral presentation · Non-archival extended abstract
Towards Neuro-Symbolic Procedural Reasoning for Long-Horizon Vision-Language-Action Manipulation
A neuro-symbolic framework that combines learned VLA control with explicit task graphs and multimodal procedural memory. The structured state guides object selection, destination grounding, subgoal dispatch, and verification across long-horizon manipulation.
- Encodes action dependencies, valid transitions, and branch conditions in explicit task graphs.
- Maintains active step, completed actions, textual context, and task-relevant visual evidence in memory.
- Studies workspace clearing and surgical-instrument handling with ordered execution and conditional decisions.
ECCV 2026 · DexHAND
Paper 02
Extended abstract · Non-archival, non-proceedings
What Matters, When? Diagnosing and Improving Conditional Visual Grounding in Visuomotor Imitation Policies
A study of conditional visual grounding in visuomotor imitation policies: the visual target required for successful control changes with manipulation phase and task state, making visually similar distractors a source of systematic failure.
- Localizes distractor-induced failures to picking and placement under controlled color and shape similarity.
- Evaluates distractor augmentation, phase-dependent attention regularization, and appearance-based visual prompting.
- Demonstrates improved robustness in simulation, on a physical UR3e, and in a pretrained VLA task.
Citation
BibTeX
If this work is useful in your research, please cite the corresponding paper.
Neuro-Symbolic Procedural Reasoning
@misc{chavan2026neurosymbolic,
title = {Towards Neuro-Symbolic Procedural Reasoning for Long-Horizon Vision-Language-Action Manipulation},
author = {Chavan, Vivek and Shi, Yahuan and Heimann, Oliver and Haninger, Kevin and Krüger, Jörg},
year = {2026},
eprint = {2609.05369},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.05369}
}
Conditional Visual Grounding
@misc{chavan2026whatmatters,
title = {What Matters, When? Diagnosing and Improving Conditional Visual Grounding in Visuomotor Imitation Policies},
author = {Chavan, Vivek and Xie, Pengtao and Shi, Yahuan and Heimann, Oliver and Haninger, Kevin and Krüger, Jörg},
year = {2026},
eprint = {2609.05376},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.05376}
}