Organizations: State Key Laboratory for Novel Software Technology, Nanjing University · Faculty of Computility Microelectronics, Shenzhen University of Advanced Technology · Amap, Alibaba Group · Istituto Italiano di Tecnologia · Shenzhen University · Xi’an Jiaotong University · Beijing University of Chemical Technology · Guangdong Provincial Key Laboratory of Computility Microelectronics
Despite the rapid progress of vision-language-action (VLA) models, the prevailing practice of predicting action chunks as discrete waypoints remains structurally misaligned with the intrinsic continuity of physical motion. This discretization arises naturally from fixed-rate robot data collection and the token-by-token prediction paradigm of large language models, but ties actions to rigid sampling rates, does not naturally support analytically consistent higher-order derivatives, and introduces quantization artifacts that hinder precise, compliant interaction. We propose Neural Implicit Action Fields (NIAF), which reformulates chunk-level action representation from discrete waypoints to continuous action functions. Using a vision-language model as a hierarchical spectral modulator over a learnable motion prior, NIAF synthesizes continuous-time action manifolds with arbitrary temporal resolution. This formulation enables analytical differentiation, allowing explicit supervision of velocity and regularization of higher-order derivative signals to promote mathematical consistency, physical plausibility, and control smoothness. Our approach achieves strong results on CALVIN and LIBERO across diverse backbones. Real-world experiments further confirm that NIAF supports stable impedance control, bridging policy-side action generation and execution-side smooth control.
Figures & tables
Figure 1: From Discrete Waypoints to Continuous Functions. Prevalent methods learn temporally discrete waypoints bound to a fixed action sampling rate. We instead model the action trajectory as a continuous time function, which offers two key advantages: 1) Resolution Independence, enabling querying at arbitrary control frequencies without interpolation artifacts; 2) Analytical Differentiability, which allows for explicit velocity supervision and jerk regularization, providing precise and smooth reference profiles required for impedance control.
Figure 2: The NIAF architecture. Instead of predicting discrete waypoints, we reformulate action generation as function regression. Operating as a hypernetwork, the MLLM serves as a hierarchical spectral modulator, transforming learnable query embeddings into modulation vectors conditioned on the multimodal context via one-step parallel decoding. These vectors dynamically reconfigure the shared meta-parameters of a SIREN. Consequently, the instantiated SIREN enables querying actions at arbitrary frequencies by simply sampling the continuous time domain τ .
Task
Method
Size
Pretraining
Tasks Completed in a Row (1000 chains)
Avg. Len ↑
1
2
3
4
5
ABCD → D
UP-VLA ( Zhang et al., 2025 )
1.3B
✓
0.962
0.921
0.879
0.842
0.812
4.42
RoboVLMs ( Li et al., 2024a )
1.6B
✓
0.967
0.930
0.899
0.865
0.826
4.49
UniVLA ( Wang et al., 2025b )
9B
✓
0.985
0.961
0.931
0.899
0.851
4.63
BEAST ( Zhou et al., 2026 )
0.77B
✗
0.981
0.962
0.930
0.893
0.848
4.61
FLOWER ( Reuss et al., 2025 )
1B
✗
0.989
0.967
0.939
0.902
0.855
4.62
Table 1: Long-horizon robotic manipulation evaluation on the CALVIN benchmark. Performance is evaluated under the ABCD → D and ABC → D settings using 1000 evaluation chains. We report the success rate of completing 1–5 tasks consecutively and the average episode length (Avg. Len). Pretraining indicates whether large-scale robot data pretraining is used. Best results are shown in bold.
Method
Size
Pretraining
Spatial
Object
Goal
Long
Average
π0 ( Black et al., 2024 )
3B
✓
96.8
98.8
95.8
85.2
94.2
π0 FAST ( Pertsch et al., 2025 )
3B
✓
96.4
96.8
88.6
60.2
85.5
OpenVLA-OFT ( Kim et al., 2025a )
7B
✓
96.2
98.2
95.6
92.0
95.5
FLOWER ( Reuss et al., 2025 )
1B
✓
97.1
96.7
95.6
93.5
95.7
BEAST ( Zhou et al., 2026 )
0.77B
✗
92.9
97.5
93.1
86.4
92.5
NIAF (ours)
0.77B
✗
98.2
100.0
98.0
95.5
97.9
Table 2: Experimental Results on the LIBERO Benchmark. Success rate (%) is reported for each task suite. NIAF is trained and evaluated separately for each suite, best results in each column are shown in bold.
Figure 3: Performance comparison of different action representations under an identical Florence-2 Large backbone on the most challenging settings.
Figure 4: Experimental Results on Real-World Robot Tasks. (a) compares different action representations under the same Florence-2 backbone. (b) evaluates portability by integrating different action heads into the π0.5 backbone.
Figure 5: Comparison of control dynamics across different methods. The top row shows joint position tracking, and the bottom row visualizes velocity and acceleration profiles. (a) & (b) Baselines (BEAST & OFT): The velocity profiles exhibit high-frequency oscillations hovering around zero, indicating disjointed stop-and-go motion and poor temporal coherence. (c) NIAF (Ours, Impedance Control): In contrast, our method produces a continuous, trend-following velocity reference (red line) that maintains consistent motion without reverting to zero-mean noise. The executed velocity (blue) aligns tightly with this analytical signal, demonstrating effective impedance tracking.
Model
Spatial
Object
Goal
Long
Avg
Qwen3-VL-FAST
97.3
97.4
96.3
90.6
95.4
Qwen3-VL-OFT
97.8
98.6
96.2
93.8
96.6
Qwen3-VL-PI
98.8
99.6
95.8
88.4
95.7
Qwen3-VL-GR00T
97.8
98.8
97.4
92.0
96.5
Qwen3-VL-NIAF
99.4
99.4
98.0
94.0
97.7
Table 3: Comparison of VLA models on LIBERO. All methods are implemented on the same Qwen3-VL backbone and trained from scratch using a single policy across four task suites. Scores are averaged over 500 trials per suite, all policies use chunking length 8 .
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Key frames of real-world evaluation rollouts. These frames visualize the successful completion of tasks by NIAF during evaluation.
Hyperparameter
LIBERO
CALVIN
Spatial
Object
Goal
Long
ABCD → D
ABC → D
Action Chunk Length
10
10
10
10
10
10
Optimizer
AdamW
AdamW
AdamW
AdamW
AdamW
AdamW
Betas
[0.9, 0.95]
[0.9, 0.95]
[0.9, 0.95]
[0.9, 0.95]
[0.9, 0.95]
[0.9, 0.95]
Learning Rate
2e-5
2e-5
2e-5
2e-5
1e-5
1e-5
Batch Size
32
32
32
32
32
32
Appendix
Table 4: Hyperparameters for simulation benchmarks.
Hyperparameter
NIAF
π0.5 -NIAF
(Item Placement, Cup Stacking)
(Towel Folding, Shape Insertion)
Action Chunk Length
50
50
Optimizer
AdamW
AdamW
Betas
[0.9, 0.95]
[0.9, 0.95]
Learning Rate
1e-5
2.5e-5
Batch Size
16
32
Appendix
Table 5: Hyperparameters for real-world experiments.
Figure 7: NIAF enables arbitrary action sampling rate. By upsampling the action chunk from 50 to 200 steps, the velocity fluctuations are reduced comparing to baselines, highlighting the smoothness of the learned manifold.
Figure 8: Water Cup Transport: Control dynamics comparison. Position (top) and velocity (bottom) profiles during transport. NIAF (pos+vel) exhibits the smoothest velocity profile, confirming that explicit velocity supervision effectively suppresses teleoperation jitter.
Figure 9: Fitting comparison: SIREN (pos+vel) vs. B-spline (pos+vel). Both models are fit to the same action chunk with position and velocity supervision. SIREN captures fine-grained local corrections while B-spline acts as a low-pass filter that underfits micro-adjustments.
Figure 10: Fitting comparison: SIREN (pos+vel+jerk) vs. B-spline (pos+vel). Adding analytical jerk regularization further smooths the velocity profile while retaining positional accuracy, demonstrating SIREN’s ability to decouple smoothness from expressiveness.
Ablation Aspect
Variant
Success Length ↑
Action Chunk Size ( H )
5
3.91
10
4.47
15
4.22
20
3.97
Weight Groups ( G )
16
4.05
32
4.20
Appendix
Table 6: Ablation Studies on CALVIN ABC → D. We evaluate the impact of key components on the average success length. The default configuration (in bold ) uses a chunk size K=10 , G=64 weight groups, and sine activation functions.
Flow-based vision-language-action (VLA) policies offer strong expressivity for action generation, but suffer from a fundamental inefficiency: multi-step inference is required to recover action structure from uninformative Gaussian noise, leading to a poor efficiency-quality trade-off under real-time constraints. We address this issue by rethinking the role of the starting point in generative action modeling. Instead of shortening the sampling trajectory, we propose CF-VLA, a coarse-to-fine two-stage formulation that restructures action generation into a coarse initialization step that constructs an action-aware starting point, followed by a single-step local refinement that corrects residual errors. Concretely, the coarse stage learns a conditional posterior over endpoint velocity to transform Gaussian noise into a structured initialization, while the fine stage performs a fixed-time refinement from this initialization. To stabilize training, we introduce a stepwise strategy that first learns a controlled coarse predictor and then performs joint optimization. Experiments on CALVIN and LIBERO show that our method establishes a strong efficiency-performance frontier under low-NFE (Number of Function Evaluations) regimes: it consistently outperforms existing NFE=2 methods, matches or surpasses the NFE=10 π0.5 baseline on several metrics, reduces action sampling latency by 75.4%, and achieves the best average real-robot success rate of 83.0%, outperforming MIP by 19.5 points and π0.5 by 4.0 points. These results suggest that structured, coarse-to-fine generation enables both strong performance and efficient inference. Our code is available at https://github.com/EmbodiedAI-RoboTron/CF-VLA.
Fan Du, Feng Yan, Jianxiong Wu +8
1Southern University of Science and Technology · 2Xi’an Jiaotong University · 3United Nova Technology +1
Vision-language-action (VLA) policies leverage pretrained vision-language backbones to achieve strong cross-task generalization. A leading design couples this backbone with a dedicated continuous action head trained via diffusion or flow matching. However, such heads rely on iterative multi-step sampling, for example 10 Euler steps in π0.5. This creates an inference bottleneck that produces stop-and-go movement in the robot and slower task completion. We introduce IMLE-VLA, which replaces the iterative action head with a single-step conditional generator trained via conditional Implicit Maximum Likelihood Estimation (cIMLE). The cIMLE objective promotes multimodal action coverage, avoiding the mode collapse of naive regression heads while eliminating multi-step sampling entirely. When IMLE-VLA is applied to π0.5, it increases inference frequency 3.67x (55 Hz vs. 15 Hz), enabling up to 11x higher action throughput. On the 40-task LIBERO benchmark, IMLE-VLA achieves the highest average success rate (98.0%) among all baselines while leading in inference frequency. Under the test-time perturbations of LIBERO-plus, IMLE-VLA retains π0.5's robustness while other baselines degrade sharply, confirming that the cIMLE head preserves generalization. Real-world experiments on a Franka Emika Panda across four tasks demonstrate smoother motion (2.2x to 3.0x lower jerk) and faster task completion, with IMLE-VLA outperforming π0.5 on every task and reducing average VLA inference time per episode by 3.9x to 6.6x. Videos and code are available at https://kianhk6.github.io/IMLE-VLA/
Kian Hosseinkhani, Qinhe Peng, George Shramko +6
Simon Fraser University · University of Pennsylvania · Alberta Machine Intelligence Institute +1
Most vision-language-action models represent future motion as fixed-rate action chunks, tying temporal resolution and prediction horizon to a fixed output budget. This pointwise representation wastes capacity on highly correlated neighboring actions, leaves temporal continuity and smoothness to be learned implicitly, and forces a tradeoff between long-horizon coverage and the local precision required for contact-rich manipulation. To address these limitations, we introduce Vela, a vision-language-action foundation model that represents future robot behavior as continuous trajectories. Vela combines a compact spline-based action representation with motion-dependent temporal support and a shared action interface for heterogeneous embodiments, allowing a fixed output budget to adapt its temporal resolution across motions. We pretrain Vela on large-scale multi-embodiment robot data and evaluate it on LIBERO-X, EBench, and two real-world long-horizon tasks, egg-cake cooking and potato shredding, obtaining promising results across simulation and physical manipulation. These results highlight the potential of continuous action representations as a foundation for future embodied foundation models. Project page and more results: https://clementine24.github.io/Vela/ .
Yifan Li, Jiaxu Wang, Dongming Wu +4
Fudan University · Shanghai Innovation Institute · The Chinese University of Hong Kong +1