Habilis-$β$: A Fast-Motion and Long-Lasting On-Device Vision-Language-Action Model

Tommoro Robotics; :; Jesoon Kang; Taegeon Park; Jisu An; Soo Min Kimm; Jaejoon Kim; Jinu Pahk

cs.ROFeb 21, 2026

Habilis- $β$ : A Fast-Motion and Long-Lasting On-Device Vision-Language-Action Model

Authors: Tommoro Robotics, :, Jesoon Kang, Taegeon Park, Jisu An, Soo Min Kimm, Jaejoon Kim, Jinu Pahk, +17 more

Organizations: Tommoro Robotics

Abstract

We introduce Habilis- $β$ , a fast-motion and long-lasting on-device vision-language-action (VLA) model designed for real-world deployment. Current VLA evaluation remains largely confined to single-trial success rates under curated resets, which fails to capture the fast-motion and long-lasting capabilities essential for practical operation. To address this, we introduce the Productivity-Reliability Plane (PRP), which evaluates performance through Tasks per Hour (TPH) and Mean Time Between Intervention (MTBI) under a continuous-run protocol that demands both high-speed execution and sustained robustness. Habilis- $β$ achieves high performance by integrating language-free pre-training on large-scale play data for robust interaction priors with post-training on cyclic task demonstrations that capture state drift across consecutive task iterations. The system further employs ESPADA for phase-adaptive motion shaping to accelerate free-space transit, utilizes rectified-flow distillation to enable high-frequency control on edge devices, and incorporates classifier-free guidance (CFG) as a deployment-time knob to dynamically balance instruction adherence and learned interaction priors. In 1-hour continuous-run evaluations, Habilis- $β$ achieves strong performance under the PRP metrics, compared to $π_{0.5}$ in both simulation and real-world environments. In simulation, Habilis- $β$ achieves 572.6 TPH and 39.2 s MTBI (vs. 120.5 TPH and 30.5 s for $π_{0.5}$ ), while in a real-world humanoid logistics workflow it achieves 124 TPH and 137.4 s MTBI (vs. 19 TPH and 46.1 s for $π_{0.5}$ ). Finally, Habilis- $β$ achieves the highest reported performance on the standard RoboTwin 2.0 leaderboard across representative tasks, validating its effectiveness in complex manipulation scenarios.

Figures & tables

Figure 1: Habilis- $\boldsymbol{\beta}$ Overview. (Left) Training Pipeline: The model is trained in three stages to achieve Fast-Motion and Long-Lasting capabilities. Stage 1 learns a robust, task-agnostic interaction prior via play data pre-training. Stage 2 post-trains on cyclic task demonstrations, utilizing Spatially Aware Downsampling (ESPADA) to compress casual free-space motions. Stage 3 distills the multi-step flow matching action expert into an efficient rectified flow model. (Right) Inference Pipeline: To enable On-Device operation, a pre-trained VLM prefix fuses multimodal observations to condition the distilled action expert. The reduced inference cost is reinvested into High-Frequency Control, using shorter action chunks for rapid closed-loop reactivity. Finally, Classifier-Free Guidance (CFG) acts as a deployment-time knob to dynamically balance instruction adherence and learned interaction priors.

Figure 2: Model Architecture. Habilis- $\beta$ ’s prefix-suffix architecture uses a pre-trained VLM prefix to process multimodal inputs, conditioning a suffix action expert that generates continuous action chunks.

Figure 3: Simulation Task Setup. Simulation tasks used in our continuous-run benchmark are (top to bottom): Dump Bin Bigbin (DBB), Place Dual Shoes (PDS), and Stack Bowls Three (SBT). Detailed task procedures are described in the Simulation Task paragraph below.

Method	Data Regime	TPH	MTBI(s)	Success Rate(%)
$\pi_{0.5}$	Normal	120.5	30.5	47.8
$\pi_{0.5}$	Play + Cyclic	207.1	49.9	69.3
GR00T N1.5	Normal	1.2	20.4	0.5
GR00T N1.5	Play + Cyclic	5.0	20.6	2.1
Habilis- $\boldsymbol{\beta}$ (w/o ESPADA)	Play + Cyclic	250.5	72.4	76.8
Habilis- $\boldsymbol{\beta}$	Play + Cyclic + ESPADA	572.6	39.2	78.5

Table 1: Simulation Deployment Performance. We report results for policies trained and evaluated on three simulated tasks: Dump Bin Bigbin , Place Dual Shoes , and Stack Bowls Three . For each task, we compute TPH, MTBI, and Success Rate under the continuous-run protocol, and report the average across the three tasks.

Figure 4: Simulation Per-Task Results. We break down the continuous-run performance metrics (TPH, MTBI, and Success Rate) for each simulation task. Habilis- $\beta$ consistently outperforms baselines across all tasks, with ESPADA significantly boosting throughput (TPH) while maintaining high success rates.

Figure 5: Simulation Productivity-Reliability Plane. We plot the deployment performance of different methods in simulation. Habilis- $\beta$ (w/o ESPADA) achieves the highest reliability (MTBI), while the full Habilis- $\beta$ dramatically increases throughput (TPH), offering a configurable trade-off for different operational needs.

Figure 6: Standard RoboTwin 2.0 Benchmark Results. Success rates are measured over 100 independent episodes with curated resets under the standard RoboTwin 2.0 benchmark. Habilis-

\beta

results are averaged across three evaluation runs to reduce seed variance.

Figure 7: RB-Y1. We used a humanoid platform in real-world experiments.

Figure 8: Real-world Task Setup. We conduct continuous sequential execution of the Dual-Bin Conveyor Packing task. The workflow involves: (1) picking an item with the right hand, (2) picking an item with the left hand, (3) waiting for the conveyor, (4) dropping the right-hand object at the correct timing, (5) dropping the left-hand object at the correct timing, and (6) immediately transitioning to the next cycle (picking with the right hand). This setup challenges the policy to maintain precise timing and coordination over long-duration runs.

Method	Data Regime	TPH	MTBI(s)	Success Rate(%)
$\pi_{0.5}$	Normal	19	46.1	19.6
$\pi_{0.5}$	Play + Cyclic	79	94.6	67.5
GR00T N1.5	Normal	33	53.7	33.0
GR00T N1.5	Play + Cyclic	63	78.2	57.7
Habilis- $\boldsymbol{\beta}$	Play + Cyclic + ESPADA	124	137.4	82.7

Table 2: Real-world Deployment Performance. We compare Habilis-

\beta

against baselines under a 1-hour continuous-run protocol. While baselines struggle with drift, Habilis-

\beta

yields a

6.53\times

improvement in productivity (TPH) and

2.98\times

reliability (MTBI) compared to the

\pi_{0.5}

baseline.

Figure 9: Real-world Productivity-Reliability Plane. We visualize the performance landscape, where the ideal system resides in the top-right quadrant (high throughput, high robustness).

Figure 10: Real-World Continuous Execution Timelines. We show 1-hour continuous-run execution logs for each method. Green segments indicate successful task completions, whereas dark blue segments indicate failed attempts that do not complete within the allowed window or are terminated by the operator, and thus trigger an intervention. Markers denote intervention categories: orange diamonds for Timeout and red triangles for Abort . Compared to baselines, Habilis- $\beta$ achieves a higher density of successes with fewer interventions, supporting its long-lasting robustness.

Explore similar work

Feb 24, 2026

Recursive Belief Vision Language Action Models

Vision-language-action models must enable agents to execute long-horizon tasks under partial observability. However, most existing approaches remain observation-driven, relying on short context windows or repeated queries to vision-language models (VLMs). This leads to loss of task progress, action repetition under perceptual aliasing, and high inference latency. While semantic grounding is important, long-horizon manipulation fundamentally requires persistent, action-conditioned state representations. Current VLAs lack such representations and exhibit limited temporal and physical reasoning, making them ill-suited for multi-stage control. This paper introduces RB-VLA, a belief-centric architecture trained with self-supervised world-model objectives that maintains a compact latent state encoding task-relevant history, dynamics, and object interactions. Queried once per task, the VLM provides high-level intent, while the belief tracks task progress and enables phase-aware, causally grounded control under partial observability without storing raw observations or scaling memory with time. The belief and intent jointly condition a diffusion policy for robust closed-loop execution. RB-VLA outperforms prior VLAs on long-horizon benchmarks, achieving 52.5 percent and 37.5 percent higher success rates on multi-stage pick-and-place and stacking tasks, respectively, compared to pi_0. It also reduces inference latency by up to five times relative to baselines and eliminates memory growth across timesteps observed in existing VLAs. Ablations show the belief module is the primary driver of performance, increasing success rates from 32.5 percent without belief to 77.5 percent with belief.

Vaidehi Bagaria, Bijo Sebastian, Nirav Kumar Patel

Department of Engineering Design, Indian Institute of Technology, Madras, Chennai, India

Vision-Language-Action Models Long-Horizon Robotic Manipulation VLMs for Robotics World Models for Robotics Efficient VLA Model Inference Belief-Space Planning

Ask PDF

Apr 7, 2026

A1: A Fully Transparent Open-Source, Adaptive and Efficient Truncated Vision-Language-Action Model

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for open-world robot manipulation, but their practical deployment is often constrained by cost: billion-scale VLM backbones and iterative diffusion/flow-based action heads incur high latency and compute, making real-time control expensive on commodity hardware. We present A1, a fully open-source and transparent VLA framework designed for low-cost, high-throughput inference without sacrificing manipulation success; Our approach leverages pretrained VLMs that provide implicit affordance priors for action generation. We release the full training stack (training code, data/data-processing pipeline, intermediate checkpoints, and evaluation scripts) to enable end-to-end reproducibility. Beyond optimizing the VLM alone, A1 targets the full inference pipeline by introducing a budget-aware adaptive inference scheme that jointly accelerates the backbone and the action head. Specifically, we monitor action consistency across intermediate VLM layers to trigger early termination, and propose Inter-Layer Truncated Flow Matching that warm-starts denoising across layers, enabling accurate actions with substantially fewer effective denoising iterations. Across simulation benchmarks (LIBERO, VLABench) and real robots (Franka, AgiBot), A1 achieves state-of-the-art success rates while significantly reducing inference cost (e.g., up to 72% lower per-episode latency for flow-matching inference and up to 76.6% backbone computation reduction with minor performance degradation). On RoboChallenge, A1 achieves an average success rate of 29.00%, outperforming baselines including pi0(28.33%), X-VLA (21.33%), and RDT-1B (15.00%).

Kaidong Zhang, Jian Zhang, Rongtao Xu +20

SYSU · Spatialtemporal AI · MBZUAI

Vision-Language-Action Models Efficient VLA Model Inference Robotic Manipulation

Ask PDF

Feb 26, 2026

Rethinking the Practicality of Vision-language-action Model: A Comprehensive Benchmark and An Improved Baseline

Vision-Language-Action (VLA) models have emerged as a generalist robotic agent. However, existing VLAs are hindered by excessive parameter scales, prohibitive pre-training requirements, and limited applicability to diverse embodiments. To improve the practicality of VLAs, we propose a comprehensive benchmark and an improved baseline. First, we propose CEBench, a new benchmark spanning diverse embodiments in both simulation and the real world with consideration of domain randomization. We collect 14.4k simulated trajectories and 1.6k real-world expert-curated trajectories to support training on CEBench. Second, using CEBench as our testbed, we study three critical aspects of VLAs' practicality and offer several key findings. Informed by these findings, we introduce LLaVA-VLA, a lightweight yet powerful VLA designed for practical deployment on consumer-grade GPUs. Architecturally, it integrates a compact VLM backbone with multi-view perception, proprioceptive tokenization, and action chunking. To eliminate reliance on costly pre-training, LLaVA-VLA adopts a two-stage training paradigm including post-training and fine-tuning. Furthermore, LLaVA-VLA extends the action space to unify navigation and manipulation. Experiments across embodiments demonstrate the capabilities of generalization and versatility of LLaVA-VLA , while real-world mobile manipulation experiments establish it as the first end-to-end VLA model for mobile manipulation. We will open-source all datasets, codes, and checkpoints upon acceptance to foster reproducibility and future research.

Wenxuan Song, Jiayi Chen, Xiaoquan Sun +12

The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. · Huazhong University of Science and Technology, Wuhan, China. · Westlake University, Hangzhou, China. +1

Vision-Language-Action Models Efficient VLA Models Mobile Manipulation

Ask PDF