IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation
:::info Stub — Full Engineering Breakdown Coming This paper has a linked code implementation and was featured on Hugging Face Papers with 14 upvotes. A full breakdown with production viability rating, implementation notes, and honest limitations is being written. Subscribe to AI Letters → :::
| Authors | Shijie Lian et al. |
| Year | 2026 |
| HF Upvotes | 14 |
| arXiv | 2605.14712 |
| Download | |
| Code | https://github.com/ZGC-EmbodyAI/IntentVLA |
Abstract
Robot imitation data are often multimodal: similar visual-language observations may be followed by different action chunks because human demonstrators act with different short-horizon intents, task phases, or recent context. Existing frame-conditioned VLA policies infer each chunk from the current observation and instruction alone, so under partial observability they may resample different intents across adjacent replanning steps, leading to inter-chunk conflict and unstable execution. We introduce IntentVLA, a history-conditioned VLA framework that encodes recent visual observations into a compact short-horizon intent representation and uses it to condition chunk generation. We further introduce AliasBench, a 12-task ambiguity-aware benchmark on RoboTwin2 with matched training data and evaluation environments that isolate short-horizon observation aliasing. Across AliasBench, SimplerEnv, LIBERO, and RoboCasa, IntentVLA improves rollout stability and outperforms strong VLA baselines
Engineering Breakdown
Plain English
This paper addresses a fundamental problem in robot learning: when training data shows the same visual scene followed by different actions (because human demonstrators had different intentions), existing VLA models fail to maintain consistent behavior across replanning steps. IntentVLA solves this by adding a history encoder that captures short-horizon intent from recent observations, then uses that intent representation to condition action generation, maintaining stability across sequential action chunks.
Key Engineering Insight
The core insight is that partial observability in robot tasks requires conditioning on latent intent state, not just current frames—by encoding recent visual history into a compact representation and threading it through the policy, you eliminate the inter-chunk conflicts that plague frame-conditioned models during replanning.
Why It Matters for Engineers
For engineers deploying robot manipulation systems, this directly solves the 'jittery' or unstable execution problem where robots make contradictory decisions on successive planning cycles even in the same scene. Most production systems today use frame-only conditioning and suffer from this aliasing problem; this approach is a practical fix that works with existing VLA architectures.
Research Context
Prior VLA work assumed single deterministic actions per observation, which breaks under real-world multimodality. This paper recognizes that robot behavior is inherently context-dependent and proposes a learnable short-horizon intent bottleneck as the missing variable. It advances the field by making VLAs robust to the aliasing problem, enabling more stable long-horizon task execution without architectural overhauls.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::