TAD: Distilling Vision-Language-Action Policies from Teacher Trajectories Alone

Yun Zhu; Da Wang

cs.AIOct 24, 2026

TAD: Distilling Vision-Language-Action Policies from Teacher Trajectories Alone

Authors: Yun Zhu, Da Wang

Organizations: Independent Researcher, China · School of Computer and Information Technology, Shanxi University, Taiyuan, China

Abstract

Online distillation of vision-language-action (VLA) policies commonly relies on teacher action scores that an inference-only interface may not expose. Trajectory-only Adversarial Distillation (\gad{}) instead turns a fixed bank of teacher joint-position features into adaptive feedback: a discriminator compares feature marginals from teacher and student rollouts and supplies rewards for on-policy learning. The student interacts with the environment without online teacher queries or environment task rewards. On ManiSkill PutOnPlateInScene25Main, \gad{} improves OpenVLA-OFT (7B) from $26.2\%$ to $76.6\%$ \successonce{} ( $+50.4$ percentage points) and from $20.3\%$ to $58.6\%$ \successend{}. It trails white-box on-policy distillation (OPD) at $81.6\%$ / $59.4\%$ by $5.0$ / $0.8$ percentage points ( $n{=}256$ evaluation trajectories per checkpoint). The tested four-epoch continuation achieves the highest \gad{} result without additional teacher data. This best run uses $5\times$ OPD's student interaction budget and $12.5\times$ its complete-rollout batch-passes. The comparison demonstrates online VLA improvement without teacher scoring while quantifying unequal interaction and optimization budgets.

Figures & tables

Panel of Figure 1 — Figure 1: Fixed teacher data, online student learning. The top path collects teacher features once. Below, the student acts in the environment, receives observations, and learns from discriminator rewards. The teacher never scores student actions during training.

Variant	Configuration	Eval success	Observation
v1	$[s_{t},a_{t}]$ 64-dim; GAIL reward $-\log(1\!-\!D)$ ; 100 $D$ updates/step; no regularization	25.4%	Accuracy approaches 0.998 within a few steps.
v2	8 $D$ updates/step; $D$ lr $1\text{e-4}$ ; label smoothing 0.1	$\sim$ 26%	Accuracy stays around 0.7–0.8 without a corresponding task gain.

Table 1: Discriminator-design diagnostics. Success is standalone deterministic success-once; these variants change several discriminator settings together.

Variant	Configuration change	Eval success	Observation
v3	v2 + environment reward 0.5; $D$ updates $8\to 20$ per step	26.6%	No substantial evaluation gain in this run.
v4	Random-noise reward; no discriminator signal	–	Sampled rollout success peaks at 42%; no standalone evaluation.
v6	State-only 8-dim qpos; AIRL logit; weight decay 0.01; four-batch replay	26.2%	Accuracy is about 0.78, but task evaluation remains near baseline.

Table 2: Reward diagnostics. Eval denotes standalone deterministic success-once; v4 has only a sampled rollout observation. v6 changes input, reward, weight decay, and replay together.

Variant	Actor configuration	Success	Observation
v6	Actor lr $2\text{e-6}$	26.2%	Low-learning-rate evaluation baseline.
v7	Actor lr $1\text{e-5}$	$\sim$ 28% *	Limited rollout improvement in the observed window.
v9	Actor lr $1\text{e-5}$ ; four update epochs	$\sim$ 28% *	Additional epochs did not improve the observed rollout level.
v8	Actor lr $5\text{e-5}$	68.8%	Standalone deterministic evaluation at step-150.

Table 3: Learning-rate and update-count diagnostics using the v6 discriminator recipe. Reported budgets and checkpoints are not matched; * denotes sampled rollout success rather than standalone evaluation.

Figure 2: Continuations from step-150. Thin lines are raw sampled success; thick lines are centered nine-point moving averages with available-point boundary windows. The gray band is the baseline v8 raw range, not a confidence interval. The red dotted line is the baseline step-150 deterministic success-once (

68.8\%

), not a branch evaluation; annotations give each branch’s independent evaluation.

Configuration	Change	Log steps	Raw success	Eval once / end (checkpoint)
Baseline v8	–	150–300	41.8–77.0%	68.8% / 46.5% (s150) 66.4% / 44.1% (s280)
Gradient penalty	gp_coef=10	196–300	48.0–77.0%	69.1% / 41.0% (s300)
Policy epochs	update_epoch=4	151–300	45.7–82.0%	76.6% / 58.6% (s300)
Env-reward mix	env_reward_coef=0.05	151–300	40.6–75.0%	69.5% / 46.5% (s300)

Table 4: Continuation results. Raw ranges cover the full available log intervals listed below. Evaluations use 256 trajectories each; checkpoint labels distinguish the baseline’s step-150/280 evaluations from the branches’ step-300 evaluations.

Method	ckpt	success-once	success-at-end	$n$
Student (SFT)	–	26.2%	20.3%	256
TAD (black-box)	step-150	68.8%	46.5%	256
TAD (black-box)	step-280	66.4%	44.1%	256
TAD + 4 $\times$ update epochs	step-300	76.6%	58.6%	256
OPD (white-box)	step-20	73.8%	53.1%	256
OPD (white-box)	step-60	81.6%	59.4%	256

Table 5: Standalone deterministic evaluation on ManiSkill PutOnPlateInScene25Main , with 256 trajectories per checkpoint. The four-update-epoch branch resumes from TAD step-150 and continues to step-300.

Figure 3: Sampled training success versus outer iteration. Thin lines are raw values; centered moving averages use 11 points for baseline TAD, 5 for OPD, and 9 for the dashed four-epoch continuation (ue4) from step-150, with available-point boundary windows. Stars mark deterministic evaluations under a different decoding protocol. The ue4 branch reaches

76.6\%

evaluation success despite overlapping the baseline’s sampled rollout range.

Figure 4: Deterministic success-once and success-at-end (

n{=}256

trajectories each). Once-to-end gaps are 22.3pp for baseline TAD at step-150, 22.2pp for OPD at step-60, and 18.0pp for TAD+ue4 at step-300.

Figure 5: One episode from the TAD step-150 deterministic evaluation, shown from left to right. The student places the object on the plate (green frame, success-once true), then re-grasps and removes it (red frame, success-at-end false).

Visuomotor Policy Learning

Visuomotor policy learning trains robots to turn camera observations, often with language and proprioception, into actions for manipulation and navigation.

See all 286 papers

Offline Semantic Guidance for Efficient Vision-Language-Action Policy Distillation
May 15, 2026Jin Shi, Brady Zhang, Yishun LuVLM Distillation Efficient VLA Model Inference
Department of Mechanical Engineering University College London London, United Kingdom · Department of Engineering Science University of Oxford Oxford, United Kingdom
ROAD-VLA: Robust Online Adaptation via Self-Distillation for Vision-Language-Action Models
Jun 24, 2026Kejing Wang, Toan Nguyen, Minh Hoang Nguyen +2Robustness of VLA Models Robot Policy Adaptation
Applied Artificial Intelligence Initiative Deakin University Geelong, Australia · Air Force Research Laboratory USA
Multi-Teacher On-Policy Distillation for Flow-Based Vision-Language-Action Models
Oct 15, 2026Lichao Ni, Yuhan Liu, Pengfei Li +1Flow-Based VLA Models Vision-Language-Action Models
Huawei Technologies Co., Ltd. · Tsinghua University

May 15, 2026J/K move · Enter open · S save

Offline Semantic Guidance for Efficient Vision-Language-Action Policy Distillation

Jin Shi, Brady Zhang, Yishun Lu

Department of Mechanical Engineering University College London London, United Kingdom · Department of Engineering Science University of Oxford Oxford, United Kingdom

VLM Distillation Efficient VLA Model Inference Visuomotor Policy Learning Efficient VLA Models Robotic Manipulation Vision-Language-Action Models+1

Open paper AskPDF arXiv

Billion-parameter Vision-Language-Action (VLA) policies have recently shown impressive performance in robotic manipulation, yet their size and inference cost remain major obstacles for real-time closed-loop control. We introduce VLA-AD, a distillation framework that uses a Vision-Language Model as an offline semantic supervisor to transfer large VLA teachers into lightweight student policies. Instead of relying only on low-level action imitation, VLA-AD augments teacher-provided 7-DoF action targets with high-level semantic guidance, including task phase anchors and multi-frame operating-direction descriptions. These auxiliary signals are used only during training: at test time, the student policy runs independently, with neither the VLA teacher nor the VLM required. We evaluate VLA-AD on three LIBERO benchmark suites. Using OpenVLA-7B as the teacher, our method produces a 158M-parameter student, yielding a

44\times

reduction in model size while matching the teacher with only a

0.27\%

average relative gap. The resulting policy runs at 12.5 Hz on an RTX 4090, achieving a

3.28\times

inference speedup over OpenVLA-7B. We further show that the same semantic distillation pipeline generalizes to a different

π_{0.5}

-4B teacher, where the student outperforms the teacher on two suites and remains within

0.53\%

on libero\_goal. Additional analysis indicates that phase-level supervision and multi-frame directional cues make the student less sensitive to noisy teacher actions, such as erroneous high-frequency gripper changes. Overall, VLA-AD demonstrates that offline semantic guidance from VLMs can substantially improve the efficiency, robustness, and deployability of VLA policy distillation.

Abstract

Visuomotor Policy Learning

Explore similar work

Offline Semantic Guidance for Efficient Vision-Language-Action Policy Distillation

ROAD-VLA: Robust Online Adaptation via Self-Distillation for Vision-Language-Action Models

Multi-Teacher On-Policy Distillation for Flow-Based Vision-Language-Action Models