Cloud-side inference expands the computational resources available to imitation-learning policies, but communication and inference latencies create a mismatch between prediction and execution. We propose RAPAC-DP, a response-aligned pending-action compensation framework for diffusion- and flow-based action generators. By conditioning on actions committed during inference, RAPAC-DP generates a full-horizon action chunk aligned with the scheduled handover time. A lightweight pending-action encoder and LoRA adapters learn this compensation while keeping the pretrained base policy frozen. We construct delay-conditioned training samples from delay-free demonstrations, requiring neither an explicit dynamics model nor additional data collected under delayed execution. In simulation benchmarks at the largest evaluated delays, RAPAC-DP retains 81.4% of its overall delay-free performance on Kinetix and achieves a mean success rate of 63.3% across three RoboMimic tasks. Across 450 physical trials on a Franka Panda manipulator, RAPAC-DP increases the high-delay mean success rate from 28.9% for RTC to 60.0% across static and conveyor-based pick-and-place tasks.
Figures & tables
Fig. 1: Schematic of observation–execution alignment under execution delay. (a) Naive inserts the complete request-aligned chunk at t+d , beginning with a^t(t) . The pending actions Pt,d change the robot state during inference, producing an observation–execution mismatch. (b) RAPAC-DP conditions on Pt,d and generates a complete action chunk beginning at the scheduled handover t+d .
Fig. 2: Overview of the RAPAC-DP framework. (a) Response-aligned training on a diffusion policy. Delay-free demonstrations provide pending-action conditions and full-horizon targets beginning at the sampled delay step. Only the pending-action encoder and LoRA adapters are updated while the base policy remains frozen. (b) Asynchronous online execution. During remote inference, the local controller executes committed pending actions and inserts the newly received action chunk into the execution queue at the scheduled handover time.
Fig. 3: Success rates under increasing execution delay on 12 Kinetix tasks. (a) Representative environments. (b) Overall performance, obtained by averaging the 12 environments equally within each seed before computing the across-seed mean. (c) Per-environment performance. RAPAC-DP keeps the compensation pathway enabled at every delay level, including d=0 . For RAPAC-DP, A2C2, and RTC, thick curves show the mean, shaded bands show the standard deviation across three seeds, and pale curves show individual seeds. Naive is evaluated with one seed and therefore has no uncertainty band.
Fig. 4: Success rates under increasing execution delay on RoboMimic: (a) Can, (b) Square, and (c) Transport. Ours denotes RAPAC-DP. For RAPAC-DP, A2C2, and Naive, thick curves and shaded bands denote the mean and sample standard deviation across three random seeds, while pale curves show individual runs. RTC reflects the success rate from 50 rollouts per condition.
Delay-range averages
Representative delays
Method
Low ( 0 – 3 )
High ( 4 – 7 )
Overall ( 0 – 7 )
d=0
d=3
d=7
Naive
77.4±2.4
28.2±1.5
52.8±1.9
89.1±2.7
62.9±3.9
13.6±0.8
+ LoRA
78.2±1.3
27.4±0.8
52.8±1.0
85.6±0.4
63.8±2.8
12.4±2.0
+ response-aligned target
61.6±2.5
70.2±1.2
65.9±1.2
87.1±1.0
68.0±2.0
55.8±1.4
+ pending-action sequence
87.8±1.4
70.2±1.8
79.0±1.6
89.3±1.3
87.3±1.3
63.3±2.3
TABLE I: RoboMimic ablation results. Rows are cumulative; values are mean success rates ± standard deviation over three seeds (%).
Fig. 5: Simulation environments and physical execution sequences. (a) RoboMimic simulation tasks: Can, Square, and Transport. (b) Physical pick-and-place tasks: Cube Grasping, Vegetable Grasping with a plastic eggplant, and Dynamic Grasping on a conveyor belt. Each row displays six chronological snapshots from an illustrative trial, with time progressing from left to right.
Fig. 6: Physical pick-and-place success rates across Cube Grasping, Vegetable Grasping, and Dynamic Grasping. RAPAC-DP (Ours) and RTC are evaluated under nominal execution latencies of 200 and 450 ms obtained through latency injection and mapped to discrete delays of d=3 and d=7 , respectively. Sync is evaluated under natural latency as a synchronous reference. Each bar summarizes 30 physical trials for a total of 450 trials; labels indicate success percentages, and error bars denote 95% confidence intervals.
While diffusion-based policies have impressive performance and expressivity, their long offline training slows down the data collection and policy deployment loop. We introduce Closed-Form Diffusion Policies, a class of training-free diffusion-based policies for imitation learning using the closed-form score derived from the demonstration dataset. We deploy CFDP with real-time inference with a mobile CPU in hardware experiments, showing it can successfully perform imitation directly from the dataset in milliseconds and with faster inference than neural diffusion policies. In experiments on imitation learning benchmarks, we show that CFDP is competitive against neural baselines that require hours of training, providing a favorable tradeoff between training time and performance. Finally, we show how closed-form diffusion policies act as a composable primitive that enables data-driven inference-time editing of pre-trained neural diffusion policies, including policy guidance and novel demonstration augmentation.
Raghav Mishra, Ian R. Manchester
Australian Center for Robotics, ARIAM Hub, and School of Aerospace, Mechanical and Mechatronic Engineering University of Sydney
Reinforcement learning in real world environments often suffers from severe performance degradation due to delayed feedback. Existing approaches typically mitigate performance degradation caused by observation delays by constructing augmented states or predicting the true states. However, these methods often overlook the inherent discrepancy between delayed state and true states induced by stochastic MDP. We theoretically prove the existence of such a discrepancy and show that it leads to the degradation of the optimal policy. To address this challenge, we propose Diffusion Guided Uncertainty Aware Delayed Policy Optimization (DUPO). Our method explicitly models the relationship between delayed state message and the current state using a diffusion model, and leverages the resulting discrepancy estimates to weight delayed policies. Extensive experiments on continuous robotic control tasks with multiple stochastic delays demonstrate that DUPO consistently outperforms existing methods and remains effective even under long and random delay scenarios.
Junqi Tu, Zejiao Liu, Fangfei Li +1
The Key Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology, Shanghai, China · The School of Mathematics, East China University of Science and Technology, Shanghai, China
Diffusion Policy (DP) excels in embodied control but suffers from high inference latency and computational cost due to multiple iterative denoising steps. Existing diffusion model (DMs) acceleration methods improve inference speed by reducing computational overhead, as DMs for image generation are computation-bound. In this work, we reveal that DP are I/O bound, and thus propose using speculative decoding, a parallel computing framework, to accelerate DP for the first time. However, it is non-trivial to address the following challenges: (1) How to preserve the denoising quality of the base model at reduced computational cost under time-varying task difficulty in embodied settings; and (2) How to dynamically allocate computation according to task difficulty in such environments. Therefore, we propose Temporal-aware Reinforcement-based Speculative Diffusion Policy (TS-DP), the first framework that enables speculative decoding for DP with temporal adaptivity. First, to handle dynamic environments where task difficulty varies over time, we distill a Transformer-based drafter to imitate the base model and replace its costly denoising calls. Second, an RL-based scheduler further adapts to the time-varying task difficulty by adjusting speculative parameters to maintain accuracy while improving efficiency. Extensive experiments show that TS-DP achieves up to a 4.17x NFE reduction and a 3.2x wall-clock speedup with a draft acceptance rate of over 85% in simulation. Real-world experiments further demonstrate a 3.8x NFE reduction while improving success rates on real-time tasks (e.g., a 23% gain on the conveyor-belt task), reaching an inference frequency of 26 Hz and enabling real-time control while maintaining comparable task performance. The code is available at https://github.com/ChildTang/TS-DP.