ODE Policy: Robot Learning via Action Dynamics, Not Actions

Jindou Jia; Jingliang Li; Kuangji Zuo; Tuo An; Yuxuan Hu; Mengfei Zhao; Bofan Lyu; Jiaqi Bai

cs.RONov 1, 2026

ODE Policy: Robot Learning via Action Dynamics, Not Actions

Authors: Jindou Jia, Jingliang Li, Kuangji Zuo, Tuo An, Yuxuan Hu, Mengfei Zhao, Bofan Lyu, Jiaqi Bai,

Organizations: MARS Lab, Nanyang Technological University

Abstract

Most robot policies, whether trained by imitation learning (IL) or reinforcement learning (RL), map observations to action values at discrete control steps and leave the continuity of physical motion to be learned implicitly. We propose ODE policy, which learns the dynamics of the action rather than the action itself. A neural ordinary differential equation (ODE) learns the vector field of the action by matching the integrated actions to the targets over short action chunks in IL and single control steps in RL. A lightweight diversity head captures multi-modal behavior by randomizing the direction of the field. At deployment, each action is integrated from the previous one, so the executed trajectory is continuous by construction. We evaluate it as a visuomotor policy, as the action head of vision-language-action (VLA) and world action models (WAMs), and as the actor of model-based RL against 14 baselines on unimanual, bimanual, dexterous, locomotion, and loco-manipulation tasks. In IL, against flow matching, it reaches 5.9 $\times$ the success, 5.4 $\times$ the generalization, and 3.9 $\times$ the mode balance after one epoch, and generates actions 10.8 $\times$ faster. It mostly moves more smoothly than human demonstrations. On 14 real-world tasks across single-arm, dual-arm, dexterous, and humanoid robots, it achieves the highest success rate. In RL, it further speeds up SuperMimic for humanoid motion tracking and smooths its actions. Efficient in learning, data, parameters, and execution, ODE policy offers a promising path to scaling robot learning.

Figures & tables

Figure 2: Direct-action policies and ODE policy. Both encode the image history and the instruction (IL) or the reference motion (RL). Left: Direct-action policies, whether a regression head, a generative head, or the actor of RL, map the embedding to action values at discrete control steps. Right: ODE policy instead learns the dynamics of the action, a vector field $\mathbf{f}_{\theta}$ , and integrates it in physical time, so that each action follows from the previous one. In RL, each chunk holds a single action.

Panel of Figure 2 — Algorithm 1 IL training pseudocode.

Panel of Figure 3 — Algorithm 2 IL inference pseudocode.

Figure 3: Diversity head for multi-modality. Demonstrations (gray) pass an obstacle on the left or the right. (A) The plain field averages them and hits the obstacle. (B) The head samples a direction

\bm{\epsilon}

with one mode per route and a rate

\bm{\lambda}_{k}

per step, and adds

\bm{\lambda}_{k}\odot\bm{\epsilon}

to the velocity. (C) The offset tilts the whole field, so each sample takes one route.

Figure 4: Learning efficiency in visuomotor imitation learning. Success rates on LIBERO, averaged over the four suites and three random seeds. (A) Success rate during training, with the standard deviation over seeds as shaded bands. The inset lists the first epoch at which each method reaches 85% (dashed line) and its ratio to ODE policy. (B) Gain of ODE policy over FM-DiT in percentage points with different numbers of demonstrations per task and (C) of parameters outside the encoder.

Figure 5: Learning efficiency in imitation learning at scale. Success rate on LIBERO of the ODE head (ours), the original flow matching head (FM), and an A2A head in (B), one seed. (A) The

\pi_{0.5}

VLA (10 episodes per task), with the FM expert initialized from

\pi_{0.5}

(solid) or randomly (dashed). (B) ImageWAM (25 episodes per task). Brackets mark the gain over the randomly initialized FM expert in (A) and over the original head in (B), and shaded bands the standard deviation over the four suites. Per-suite results are given in Appendix B.4 and B.5 .

Figure 6: Learning efficiency in reinforcement learning. The ODE actor vs. the direct-action actor of SuperMimic and PPO on tracking the 40 LAFAN1 clips with a G1 humanoid, over wall-clock time on one GPU. (A) Gain of the ODE actor in the training return and (B) in the normalized evaluation return on all clips, where the steps after a fall count as zero, both from the means over three random seeds. (C) Acceleration of the executed actions (lower is smoother), mean and standard deviation over three random seeds. At 0.4 hours, the direct-action actor still barely moves.

Figure 7: Inference cost and action smoothness. Results on LIBERO. (A) Inference time per chunk of 8 actions vs. success rate at epoch 100, where the shaded region exceeds the 400 ms budget of 8 control steps at 20 Hz. ODE policy is shown with RK4, RK2, and Euler (dotted line). Error bars show the standard deviation of the time over successful episodes of one seed and of the success rate over three random seeds. (B) RMS jerk of the end-effector position and (C) control steps to complete a task over the successful episodes at epochs 20, 50, and 100 (hatching), with the standard error over the episodes of three random seeds. The dashed line marks ODE policy at epoch 50, and arrows give the ratio of each baseline to it (lower is better).

Figure 8: Inference cost and action smoothness at scale. Results on LIBERO with one seed. (A) Inference time per chunk of 8 actions in

\pi_{0.5}

at epoch 3, for the action head and for the full model, on one A100 GPU with batch size 1. (B) Mean change of the normalized position command inside chunks and at the first step of a new chunk in

\pi_{0.5}

. (C) Inference time per chunk in ImageWAM with cached text features, for the model and for the full local pipeline. Hatched bars mark randomly initialized action heads.

Figure 9: Generalization to unseen perturbations. Zero-shot success rate on LIBERO-plus over all perturbation types. (A) Visuomotor policies, pooled over the four suites, with the standard deviation over three random seeds as shaded bands. (B)

\pi_{0.5}

and (C) ImageWAM on a shared set of 2008 instances, one seed, with the standard deviation over the four suites as shaded bands. Dashed lines mark randomly initialized action heads, and the legend of (B) gives the size of each head. Results per perturbation type are given in Fig. S1 and Table S4 .

Figure 10: Multi-modal tests on a 2D navigation task with four feasible routes. (A) The agent moves from the start point through one of four gaps to the target point. (B) Success rate and (C) mode balance $\gamma$ over training, mean and standard deviation over three random seeds with 100 rollouts each. The arrow marks the ratio of $\gamma$ between ODE policy and FM-DiT after one epoch.

Figure 11: Tasks on a real single-arm robot. ODE policy on a Franka Research 3 arm, one row per task. (A) Putting a mango into a box and closing the lid. (B) Hanging a key on a key holder. (C) Folding a sock and putting it into a drawer. (D) Inserting a block into a hole of a board. On the right, success rates over 20 trials, with failed trials hatched and arrows from the best baseline.

Figure 12: Four tasks on a real dual-arm robot. ODE policy on the Galaxea R1 Lite, one row per task. (A) Inserting a poppy or a dahlia into the narrow opening of a vase (dashed circle and zoom). (B) Putting a bowl into a microwave, pushing it in, and closing the door. (C) Inserting two flowers one after the other. (D) Either arm puts an apple into the drawer and the other closes it, two frames per mode. On the right, success rates over 20 trials per flower or task, with failed trials hatched and arrows from the best baseline. In (D), the share of trials in which each arm goes for the apple is shown next to it, with the dashed line at the share in the demonstrations.

Figure 13: Tasks on a real robot with a dexterous hand. ODE policy on the Sharpa North robot with a Sharpa hand, one row per task. (A) Picking up a can and placing it on a plate. (B-D) Tasks learned only from public demonstrations of another lab ( Sharpa Robotics, 2026 ) (left) and deployed in our lab (right), (B) reorienting a fallen cup upright, (C) pouring pills from one cup into another with two hands, and (D) stacking one cube on another. On the right, success rates over 20 trials, with failed trials hatched and arrows from the best baseline.

Figure 14: Loco-manipulation tasks on a real humanoid. ODE policy as the action head of GR00T N1.7 on a Unitree G1 with Inspire DFX dexterous hands, one row per task. (A) Walking to a toy on the floor, squatting to grasp it, standing up, and placing it into a box on the sofa. (B) Grasping a rice ball from a table, turning and walking to a microwave, putting the rice ball inside, and closing the door. On the right, success rates over 20 trials, with failed trials hatched and arrows from the baseline.

Figure 15: Inference cost and smoothness on real robots. The close-box, microwave, and can tasks (top to bottom). (A) Inference time per chunk on one RTX 5080, RTX 5090, and RTX 4090, in log scale. (B) Acceleration of the joint commands at chunk boundaries and inside chunks. (C) RMS jerk of the measured joint motion. Mean and standard deviation over episodes (successful ones on the single-arm robot), with the demonstrations as dashed lines.

Figure 16: Ablations of the chunk length and the integration solver. Results of ODE policy on LIBERO, averaged over the four suites. (A) Success rate with chunk length

H

at epochs 1, 10, 50, and 100. (B) Inference time per action at epoch 100 on one idle GPU, with the default

H=8

in gray. (C) Success rate during training with three solvers, with FM-DiT and A2A-DiT as references and an inset on the solvers. (D) Training time to epoch 100 (one RTX 4090 per run, eight runs in parallel). Mean and standard deviation over three random seeds (two for

H=16

, as training with the third diverged).

Figure 17: Ablations of the contractive perturbation, seamless chunking, and proprioception. Results of ODE policy on LIBERO, averaged over the four suites (10 episodes per task). (A) Without the contractive perturbation and with three magnitudes of it. (B, C) Seamless chunking vs. restarting each chunk from the measured state, success rate and boundary ratio (command change at the first step of a chunk divided by that inside chunks). (D) Proprioception fed to the vector field, to both the condition and the field, or to the condition. Arrows in (A) and (B) mark the gap after one epoch. Mean and standard deviation over three random seeds, with an inset in (C) on the largest standard deviation.

Figure 18: Ablation on the number of components. (A) Distinct groups of components used by the diversity head at the start of the 2D navigation task and per state on LIBERO, with the four routes as dashed line. (B) Mode balance

\gamma

on the navigation task at epoch 20. (C) Success rate on LIBERO at epochs 1 and 20 (10 episodes per task). Mean and standard deviation over three random seeds, with the default

K=12

in gray.

Figure 19: Two potential uses of the continuous-time field. (A, B) Learning from data without action labels. All policies are trained on the measured poses of LIBERO only and evaluated at epoch 100, and ODE policy with PD

+

velocity feeds the velocity of its field to the PD controller. (A) Success rate on five time-critical LIBERO-Long tasks and averaged over LIBERO-Long and LIBERO-Spatial, and (B) steps of successful Long episodes. (C) Arbitrary control rates, by changing the integration of one trained model at deployment, with substeps within each control step at 20 Hz or a smaller step sending every intermediate point at 40 or 80 Hz. Average success rate over the four suites. One seed with 10 episodes per task. In (A), arrows give the gain of ODE policy with PD

+

velocity over FM-DiT. Whiskers in (B) and (C) show the standard deviation across tasks.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Micro call	Micro encoder	Micro head	Call	Action	SR	Params
ODE-MLP (ours)	8.82	1.22	7.6	9.69	1.25	96.7	35.6
FM-DiT	42.1	1.28	40.8	41.3	5.31	93.3	35.1
FM-UNet	58.8	1.22	57.5	60.1	7.7	86.7	36.0
A2A-DiT	3.68	1.24	2.44	4.79	0.617	96.7	35.1
DDPM-DiT	235	1.24	233	237	30.5	90.0	34.7
DDPM-UNet	335	1.25	333	341	43.5	93.3	36.0

Appendix

Table S1 : Inference time (ms) on one idle RTX 4090 with the epoch-100 checkpoints. Micro timings repeat a fixed input 200 times, and call and action timings are measured in closed loop on the first three tasks of LIBERO-Goal (10 episodes each), where SR is the success rate (%). Params counts all parameters (M).

Figure S1 : Zero-shot success rate of the visuomotor policies per LIBERO-plus perturbation type at epoch 100, pooled over the four suites and three random seeds. The number above each group is ODE policy, and error bars show the standard error over the pooled instances.

Figure S2 : Rollouts on the 2D navigation task at epochs 1, 3, 5, 10, and 20, 100 per panel (stars mark the start, circles the goal, and colors the gap taken, with failed rollouts in gray). One of the three random seeds.

Method	Spatial	Object	Goal	Long	LIBERO	LIBERO-plus
ODE (ours)	96.0	100.0	96.0	92.0	96.0	71.5
FM	90.0	91.0	90.0	81.0	88.0	69.2
FM-R	83.0	89.0	71.0	59.0	75.5	52.6

Appendix

Table S2 : The

\pi_{0.5}

VLA per LIBERO suite, success rate (%) at epoch 1, one seed, with 10 episodes per task. FM starts from the pretrained action expert of

\pi_{0.5}

, and -R marks a randomly initialized one. LIBERO-plus is the shared set of 2008 instances.

Figure S3 : ImageWAM per suite on LIBERO (A-D) and on the shared LIBERO-plus set (E-H) , with the ODE head (ours), an A2A head, and the original flow matching head (FM). Whiskers show the standard deviation over the ten tasks of each suite.

Setting	FM (original)	A2A	ODE (ours)
Action generation	10 flow matching steps	1 flow step from past actions	1 condition read and 12 Euler steps
Predicted / executed actions per chunk	16 / 12	16 / 12	12 / 12
Action space	controller command	controller command	desired action, converted to the controller command
Proprioception	condition token	condition token	input of the vector field
Gripper	part of the action flow	part of the action flow	separate head
Trainable parameters	4517.6M	4517.6M	4518.2M

Appendix

Table S3 : ImageWAM settings that differ between the action heads. All heads use AdamW with a learning rate of

10^{-4}

, a warm-up of half an epoch, and a cosine decay to

10^{-6}

, the current images of two cameras at

224\times 224

without image history, and no EMA, and each epoch is one pass over the same data. Each LIBERO checkpoint is evaluated with 25 episodes per task.

	Shared set (2008)			ODE (ours), full set (10030)
Perturbation	FM	A2A	ODE (ours)	Spatial	Object	Goal	Long	All
Camera	78.7 $\pm$ 29.5	82.9 $\pm$ 24.6	91.1 $\pm$ 15.2	92.6 $\pm$ 9.5	99.2 $\pm$ 1.0	92.2 $\pm$ 12.2	75.7 $\pm$ 21.2	89.7 $\pm$ 15.2
Robot initial pose	52.5 $\pm$ 26.7	61.6 $\pm$ 24.6	71.7 $\pm$ 22.8	77.1 $\pm$ 11.8	63.3 $\pm$ 11.4	67.0 $\pm$ 17.0	68.7 $\pm$ 13.3	68.8 $\pm$ 14.2
Language	89.2 $\pm$ 17.0	92.1 $\pm$ 14.8	92.1 $\pm$ 14.2	96.7 $\pm$ 9.8	93.8 $\pm$ 11.1	89.0 $\pm$ 10.6	88.8 $\pm$ 21.1	92.0 $\pm$ 13.6
Light	97.4 $\pm$ 5.9	95.6 $\pm$ 16.2	98.7 $\pm$ 3.7	99.7 $\pm$ 0.9	100.0 $\pm$ 0.0	94.6 $\pm$ 9.5	96.4 $\pm$ 4.4	97.7 $\pm$ 5.2
Background	88.5 $\pm$ 16.6	95.1 $\pm$ 12.7	93.8 $\pm$ 16.5	95.3 $\pm$ 15.6	96.8 $\pm$ 4.9	94.0 $\pm$ 12.6	91.7 $\pm$ 11.0	94.3 $\pm$ 11.4

Appendix

Table S4 : ImageWAM on LIBERO-plus per perturbation type, success rate (%) with the standard deviation over tasks. Left, the shared set of 2008 instances at the best checkpoint of each head (epoch 8 for FM and A2A, epoch 10 for ODE). Right, ODE policy at epoch 10 on the full set of 10030 instances, per suite.

	Demonstrations	ODE policy (ours)	FM-DiT	DDIM-UNet
Time per chunk (ms)	–	10.2	35.0	51.8
95th percentile and maximum time (ms)	–	12.1 and 65.8	36.6 and 49.5	53.6 and 65.2
Acceleration at chunk boundaries (rad/s 2 )	1.04	$\mathbf{0.29\pm 0.06}$	$6.40\pm 0.82$	$10.9\pm 4.5$
Acceleration inside chunks (rad/s 2 )	1.09	$\mathbf{0.34\pm 0.06}$	$3.15\pm 0.10$	$4.02\pm 2.30$
Boundary to inside ratio	0.96	0.86	2.03	2.97
Arm jerk (rad/s 3 )	10.3	$\mathbf{6.1\pm 0.6}$	$19.6\pm 1.7$	$18.1\pm 1.4$

Appendix

Table S5 : Inference cost and smoothness on the real cup task. Inference is measured on one RTX 4090 over all chunks of all trials. Smoothness is computed over the first 12 s after the policy takes over, as the mean

\pm

standard deviation over trials. The command acceleration is taken over the 14 arm joints, and the jerk of the measured motion over the 14 arm joints and the 22 joints of the right hand. The demonstrations are the 50 training demonstrations from SharpaDex, cut into chunks of 8 steps.

	Demonstrations	ODE policy (ours)	FM-DiT
Time per chunk (ms)	–	10.3	36.3
95th percentile and maximum time (ms)	–	12.7 and 30.5	42.6 and 89.4
Acceleration at chunk boundaries (rad/s 2 )	1.57	$\mathbf{1.14\pm 0.15}$	$6.12\pm 0.58$
Acceleration inside chunks (rad/s 2 )	1.27	$\mathbf{0.78\pm 0.07}$	$2.83\pm 0.20$
Boundary to inside ratio	1.24	1.45	2.17
Arm jerk (rad/s 3 )	12.8	$\mathbf{9.8\pm 0.6}$	$19.6\pm 1.5$

Appendix

Table S6 : Inference cost and smoothness on the real pill task. Inference is measured on one RTX 4090 over all chunks of all trials. Smoothness is computed over the first 12 s after the policy takes over, as the mean

\pm

standard deviation over the 17 successful trials of ODE policy and all 20 trials of FM-DiT. The command acceleration is taken over the 14 arm joints, and the jerk of the measured motion over the 14 arm joints and the 22 joints of the right hand. The demonstrations are the 50 training demonstrations from SharpaDex, cut into chunks of 8 steps.

Figure S4 : Returns in reinforcement learning. The ODE actor of SuperMimic vs. PPO on whole-body tracking of the 40 LAFAN1 clips, over wall-clock time on one GPU. (A) Training return. (B) Normalized return on all 40 clips with the deterministic policy, where the steps after a fall count as zero. Lines and shaded bands show the mean and standard deviation over three random seeds.

Method	Vector field	Ep. 1	Ep. 10	Ep. 100	it/s	ms/chunk
ODE-MLP (ours)	FiLM-MLP	79.2	93.9	93.8	17.1	11.0
ODE-DiT (ours)	AdaLN Transformer	77.2	89.6	91.2	3.5	51.9
FM-DiT	–	13.5	80.1	85.4	31.6	61.4
A2A-DiT	–	42.2	86.9	85.6	23.3	5.8

Appendix

Table S7 : Ablation on the vector field architecture, with the DiT baselines as references. Both ODE variants use RK4 with

H=8

. Success rate (%) on LIBERO, averaged over the four suites (mean over three random seeds, except ODE-DiT with one seed, each suite with 10 tasks

\times

20 episodes), training speed (iterations per second, batch size 32), and inference time per chunk of 8 actions (ms, shared evaluation GPUs). The two ODE variants have matched parameters outside the encoder.

Inference	Epoch 1	Epoch 5	Epoch 10	Epoch 20
No offset	85.3 $\pm$ 1.9 (98/81/91/72)	95.0 $\pm$ 1.0 (99/97/96/87)	92.8 $\pm$ 1.0 (99/94/93/85)	93.1 $\pm$ 1.2 (99/97/93/84)
Component mean (ours)	79.8 $\pm$ 1.6 (98/78/83/60)	94.8 $\pm$ 0.4 (100/97/96/87)	94.3 $\pm$ 1.3 (99/98/95/86)	94.2 $\pm$ 1.0 (99/95/95/87)
Full sampling	52.1 $\pm$ 0.1 (79/50/52/28)	90.8 $\pm$ 0.0 (96/94/90/82)	93.7 $\pm$ 1.1 (98/93/96/87)	94.8 $\pm$ 0.4 (99/99/94/87)

Appendix

Table S8 : The diversity head at inference, for the same models. Without the offset (

\bm{\epsilon}=\mathbf{0}

, the plain ODE), with the mean of the drawn component, and with full sampling within the component. Success rate (%) averaged over the four LIBERO suites, mean

\pm

std over three random seeds, with the per-suite means in parentheses (Object / Spatial / Goal / Long), and 10 episodes per task.

Method	Ep.	SR	In-chunk	Cross-chunk	Jerk	Steps	ms/action	ms/call
ODE-MLP (ours)	20	92.9 $\pm$ 0.5	$\mathbf{0.01871}$ $\pm$ 0.0051	$\mathbf{0.02855}$ $\pm$ 0.0059	$\mathbf{5.257}$ $\pm$ 2.1	$\mathbf{146.5}$ $\pm$ 71	$\mathbf{1.43}$ $\pm$ 0.079	$\mathbf{11.11}$ $\pm$ 0.55
	50	93.8 $\pm$ 0.3	$\mathbf{0.0196}$ $\pm$ 0.0053	$\mathbf{0.0257}$ $\pm$ 0.0054	$\mathbf{5.577}$ $\pm$ 2.3	$\mathbf{146}$ $\pm$ 70	$\mathbf{1.432}$ $\pm$ 0.072	$\mathbf{11.15}$ $\pm$ 0.56
	100	93.8 $\pm$ 1.3	$\mathbf{0.02017}$ $\pm$ 0.0054	$\mathbf{0.02453}$ $\pm$ 0.0053	$\mathbf{5.736}$ $\pm$ 2.3	$147.2$ $\pm$ 72	$\mathbf{1.414}$ $\pm$ 0.069	$\mathbf{11}$ $\pm$ 0.47
ODE-DiT (ours)	20	89.4	$0.01914$ $\pm$ 0.005	$0.02893$ $\pm$ 0.0057	$5.276$ $\pm$ 2.1	$148.5$ $\pm$ 73	$7.192$ $\pm$ 0.23	$55.97$ $\pm$ 1.3
	50	90.8	$0.02058$ $\pm$ 0.0052	$0.02596$ $\pm$ 0.0059	$5.734$ $\pm$ 2.3	$146.6$ $\pm$ 73	$7.155$ $\pm$ 0.26	$55.63$ $\pm$ 1.5
	100	91.2	$0.02133$ $\pm$ 0.0053	$0.02516$ $\pm$ 0.0057	$5.942$ $\pm$ 2.3	$\mathbf{144.5}$ $\pm$ 69	$6.672$ $\pm$ 0.19	$51.91$ $\pm$ 0.92

Appendix

Table S9 : Ablation on the vector field architecture. Both ODE variants use RK4 with

H=8

. Execution quality and inference time on LIBERO, over the successful episodes of all suites and seeds (mean

\pm

standard deviation over episodes). SR is the mean and standard deviation of the success rate over three random seeds. In-chunk and cross-chunk denote the mean change of the command between consecutive steps inside a chunk and at the first step of a new chunk. Jerk is the RMS jerk of the measured end-effector position (m/s 3 ). Steps is the number of control steps of a successful episode, and time is the inference time per action and per call (ms) of one seed, measured on shared evaluation GPUs. ODE-DiT is trained with one seed.

Robot Policy Learning

Robot policy learning develops controllers that map sensory observations, task goals, and interaction history to actions for embodied robots.

See all 383 papers

Jul 31, 2026

Temporal Policy: History-Initialized Action Generation for Robotic Learning from Demonstration

By relying on independent couplings from uninformative Gaussian priors, standard diffusion and flow matching models are forced to learn complex, high-cost vector fields to reach the physical action space. Generative models excel at capturing multimodal behaviors for robotic Learning from Demonstration (LfD), but often suffer from high inference cost. This paper introduces Temporal Policy, a generative framework based on stochastic interpolants that formulates action generation as a temporally coupled transport problem. By initializing the generative flow at the robot's recent history, we explicitly couple past states to future action sequences. This data-dependent coupling reduces transport cost and produces straight vector fields. We validate Temporal Policy across visuomotor simulation benchmarks and on a physical Barrett WAM 2x 7DoF teleoperation platform. Our approach reduces transport costs by nearly an order of magnitude compared to noise-initialized baselines, achieving a 19.1 ms inference latency on a single NVIDIA RTX 4080. Crucially, these geometric and computational efficiencies are achieved while matching the success rates of state-of-the-art baselines. This simplified transport geometry bypasses the computational bottleneck of independent Gaussian priors, helping enable high-frequency, closed-loop control. The code is publicly available at https://github.com/dmiller12/TemporalPolicy.

Dylan Miller, Martin Jagersand

Department of Computing Science, University of Alberta, Canada

Flow Matching Robot Policy Learning Robot Skill Learning Visuomotor Policy Learning Stochastic Interpolants Robot Imitation Learning+1

Ask PDF

Oct 1, 2026

TOAST: Stochastic Robot Action Tokenization for Autoregressive Vision-Language-Action Models

Autoregressive Vision-Language-Action models often represent continuous robot actions as discrete token sequences, enabling action prediction with standard next-token objectives. FAST has substantially improved this representation by compactly encoding action containing diverse temporal frequencies into relatively few tokens. However, while such compression reduces the number of action tokens required for autoregressive prediction, it does not necessarily improve the efficiency of policy learning from limited demonstrations. In particular, FAST typically assigns a single deterministic tokenization to each quantized action sequence, although multiple token sequences can represent and decode to the same robot motion. We investigate whether exploiting this representational redundancy can improve policy learning. In this paper, we propose TOkenization of Action sequences with STochastic sampling (TOAST), a stochastic action tokenization method that samples alternative tokenizations of the same quantized action sequence during policy training. This diversifies the discrete supervision while preserving the underlying robot action and requires no additional demonstrations. Experiments on LIBERO show that TOAST consistently improves over its deterministic counterpart, with the improvement increasing as training data decreases, achieving a 6.8 point gain in success rate when only 1/16 of training data is available. Across four real-robot manipulation tasks, TOAST further improves mean success rate by 15.8 points over the deterministic counterpart. These results demonstrate the effectiveness of stochastic action tokenization for autoregressive robot policy learning, particularly when training data are limited.

Keisuke Shirai, Tomohiro Motoda, Hanbit Oh +5

AIST

Data-Efficient Robot Learning Vision-Language-Action Models

Ask PDF

Sep 16, 2026

Decoupling Vision, Language, and Action for Efficient Multi-Task Robot Policies

Vision-Language-Action (VLA) policies commonly run Vision-Language Model (VLM) backbones with billions of parameters at every policy inference, which costs latency and energy. We revisit a decoupled alternative for multi-task manipulation: separate vision and language encoders whose representations condition a compact action head. We run a standardized comparison that varies the vision encoder, the language encoder, and the action head while holding the demonstrations, the training-step budget, the tasks, the evaluation protocol, and the measurement platform fixed, against seven VLA baselines. The resulting Decoupled Embodiment Model (DEM) combines a fine-tuned DINOv3 vision encoder, a frozen NeoBERT language encoder, and a MeanFlow head that generates an action chunk in one forward pass. On 18 RoboCasa tasks evaluated with held-out instruction paraphrases and randomized scenes, DEM reaches 55.6% mean success against 56.9% for GR00T N1.7 and 54.6% for

π_{0.5}

, and on three real-robot tasks it reaches 66.0% against 68.0% for GR00T N1.7. On the same workstation, DEM needs 6.1,ms per policy forward pass, a maximum throughput of 162.7 policy calls per second, and draws an estimated 2.07,J of GPU energy per call, eight to seventeen times the throughput and six to fifteen times less energy than these VLM-backbone policies. Within this trained-task regime, DEM sits on the observed success--latency--energy frontier and provides a strong, efficient baseline for language-conditioned robot skills.

Xiatao Sun, Chen Liang, Ziyao Zeng +5

Department of Computer Science, Yale University, New Haven, CT, USA. · Peking University, Beijing, China. · Digients, Singapore.

Visuomotor Policy Learning Efficient VLA Models Robotic Manipulation Vision-Language-Action Models VLA Model Training

Ask PDF

Abstract

Robot Policy Learning

Explore similar work

Temporal Policy: History-Initialized Action Generation for Robotic Learning from Demonstration

TOAST: Stochastic Robot Action Tokenization for Autoregressive Vision-Language-Action Models

Decoupling Vision, Language, and Action for Efficient Multi-Task Robot Policies