CAPEX: Efficiently Distilling Foundation Model Behavior into Deployable Robot Policies through Experience-Adaptive Reasoning

Shivam Aarya; Zhang Xi-Jia; Chengyue Huang; Junhyun Kim; Huishu Xue; Hrishit Leen; Roman Yakunin; Animesh Garg

cs.ROSep 26, 2026

CAPEX: Efficiently Distilling Foundation Model Behavior into Deployable Robot Policies through Experience-Adaptive Reasoning

Authors: Shivam Aarya, Zhang Xi-Jia, Chengyue Huang, Junhyun Kim, Huishu Xue, Hrishit Leen, Roman Yakunin, Animesh Garg, +1 more

Organizations: School of Interactive Computing Georgia Institute of Technology Atlanta, GA 30332, USA

Abstract

Robot learning has largely relied on human-teleoperated demonstrations to acquire effective learnable behaviors. However, human-operated data collection processes can be unintuitive, difficult to scale, and inherently asynchronous. We explore an alternative: distilling physical behavior from general-purpose multimodal foundation models into deployable robot policies by using the foundation model itself as an autonomous demonstrator. While sufficiently capable models can generate successful zero-shot manipulation trajectories, repeatedly invoking them during physical execution is slow and expensive, limiting their utility as scalable data generators. As a solution, we introduce CAPEX, an experience-conditioned demonstration collection framework that uses execution experience from previous attempts to adapt how frequently the foundation model must observe, reason, and replan. We evaluate across RoboCasa tasks and on physical Franka and bimanual YAM-arm platforms, measuring task success, model calls, token usage, collection time, and cost. We further train Diffusion Policy and ACT on matched sets of human-teleoperated and foundation-model-generated demonstrations to evaluate the downstream learning value of autonomously collected data. We find that CAPEX increases the number of successful demonstrations by 4.3x while reducing the cost per successful demonstration by 80%. Policies trained on CAPEX-generated data approach the performance of those trained on matched human demonstrations; with longer training, this gap largely closes for policies trained from scratch. These results suggest that foundation models can serve as scalable sources of reusable robot experience. Project page: https://capex-paper.github.io/

Figures & tables

Figure 1: CAPEX overview. All arms start from the initial state $s_{e}$ of a human demonstration (CloseFridge shown). (a) Teleoperation yields $\mathcal{D}_{H}$ . (b) The base arm queries a VLM once per waypoint. (c) CAPEX ’s VLM proposes a multi-waypoint plan with stated confidences $q_{k}$ (circles); the robot executes it until a later waypoint’s calibrated probability $\hat{p}_{k}$ (bars) falls below $\tau$ or execution stops, then queries the VLM again. (d) Before each episode, all past attempts are summarized into $\Gamma_{e}$ : calibrator weights $\mathbf{w}_{e}$ , which set how much of a plan runs, and a VLM context $\mathcal{M}_{e}$ , which shapes the plan with the robot’s own successes $\mathcal{R}_{e}$ and how reliably each task phase has executed $\mathcal{B}_{e}$ . Student policies are trained on $\mathcal{D}_{\text{ours}}$ and $\mathcal{D}_{H}$ . Dashed arrows cross episodes.

Figure 2: Autonomous collection success of VLM(Base) and VLM + CAPEX on the same 20 recreated human-demonstration starts per task. Left: GPT-6 Astra, pooled by RoboCasa task family (120, 100, 40, 80 and 20 starts). Right: all 18 tasks for both backbones. Whiskers are 95% Wilson intervals. Astra + CAPEX outperforms Astra(Base) in 14 of the 18 tasks, 11 of them significantly (

p<0.05

, McNemar).

Method	Success $\uparrow$ (95% CI)	$/ Success$ \downarrow$	Requests / Success $\downarrow$	Min. / Success $\downarrow$
Astra (Base)	45/360 12.5% (9.5–16.3)	9.15	244	39.7
Astra + CAPEX	193/360 53.6% (48.4–58.7)	1.83	37	6.6
Qwen (Base)	0/360 0.0% (0.0–1.1)	—	$\infty$	$\infty$
Qwen + CAPEX	5/350 1.4% (0.6–3.3)	—	4,819	2,332

Table 1: Autonomous collection over the same starts for both backbones. Success is over all scorable starts with 95% Wilson intervals. Cost, model requests and wall-clock time include every attempt and are divided by the number of successful demonstrations;

\infty

marks zero successes and — marks no API price (Qwen runs locally).

\Delta

: CAPEX minus Base (percentage points for success). Bold: better within each backbone.

Figure 3: Learning from experience (GPT-6 Astra). (a) Success on the same starts with memory and with memory cleared, split by whether the memory already holds a successful execution; whiskers are 95% Wilson intervals. (b, c) Waypoint confidence before and after calibration, over the same 8,458 executed waypoints in ten equal-width confidence bins: for each bin, the mean predicted confidence and the actual share of waypoints that reached their target. Before calibration, the VLM never predicts below 0.5 and its predictions sit well above the actual reach rate; after calibration the two lines nearly coincide (ECE 0.24

\rightarrow

0.05, AUROC 0.57

\rightarrow

0.75).

	DP Fine-Tuned		DP From Scratch		ACT
Task (Demos)	Ours	Human	Ours	Human	Ours	Human
Close Fridge (50)	0.98 $\pm$ 0.02	0.89 $\pm$ 0.05	0.79 $\pm$ 0.04	0.97 $\pm$ 0.01	0.57 $\pm$ 0.11	0.93 $\pm$ 0.01
Turn On Microwave (46)	0.75 $\pm$ 0.05	0.85 $\pm$ 0.08	0.27 $\pm$ 0.09	0.50 $\pm$ 0.05	0.33 $\pm$ 0.17	0.33 $\pm$ 0.08
PnP Counter $\rightarrow$ Stove (50)	0.25 $\pm$ 0.06	0.29 $\pm$ 0.04	0.07 $\pm$ 0.01	0.01 $\pm$ 0.01	0.29 $\pm$ 0.05	0.09 $\pm$ 0.03
Close Toaster Oven (50)	0.81 $\pm$ 0.06	0.77 $\pm$ 0.06	0.51 $\pm$ 0.07	0.54 $\pm$ 0.09	0.38 $\pm$ 0.07	0.61 $\pm$ 0.11
PnP Toaster $\rightarrow$ Counter (33)	0.19 $\pm$ 0.05	0.64 $\pm$ 0.05	0.05 $\pm$ 0.06	0.06 $\pm$ 0.02	0.09 $\pm$ 0.03	0.19 $\pm$ 0.08

Table 2: Downstream policy success when trained on CAPEX demonstrations (Ours) or human teleoperation from the same initial states, with equal dataset sizes for each task. Values are mean

\pm

standard deviation over three training seeds, with 50 evaluation rollouts per seed after 20k training steps. Bold indicates the higher mean within each task–policy pair.

Figure 4: Downstream success as a function of training budget, averaged over the five tasks of the extended-budget study. From left to right: Diffusion Policy initialized from the RoboCasa365 checkpoint, Diffusion Policy trained from scratch, and ACT. The gap between CAPEX and human demonstrations largely closes with additional optimization for policies trained from scratch.

YAM: Autonomous Collection
	Success		$/ Success$ \downarrow$
Task	Base	CAPEX	Base	CAPEX
Bread + Corn	12/20	20/21	1.10	0.43
Bowl Handover	6/20	20/21	2.73	0.88
Overall	18/40	40/42	1.64	0.65

Table 3: Physical-robot evaluation. Left: autonomous demonstration collection on the bimanual YAM; cost is API cost per successful demonstration. Right: downstream policy success on the Franka Panda; each entry reports successes over 10 physical trials.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

$j$	Feature	$\phi_{k}[j]$
0	bias	$1$
1	stated log-odds	$\operatorname{logit}(q_{k})$
2	gripper closed after waypoint $k$	$\{0,1\}$
3	gripper phase (gripper commands so far)	$\min(\gamma_{k},4)/4$
4	leg length	$\lVert x_{k}-x_{k-1}\rVert/10$
5	vertical leg	$(x_{k}-x_{k-1})_{z}/10$

Appendix

Table 4: Calibrator features for waypoint

k

of a plan. Positions in cm in the robot-base frame;

x_{0}

is the current tip position and

x^{\mathrm{loc}}

the most recent locate_pixel point of the episode.

VLM	GPT-6 Astra (Responses API), reasoning effort medium, at most 8,000 output tokens per request
Open-weight VLM	Qwen3.8-27B with FP8 weights served by vLLM on one A40; temperature 1.0, top- $p$ 0.95, top- $k$ 20; 32k-token context
Plans	at most 24 waypoints; $\tau=0.5$ ; the first waypoint always runs
Episode limits	horizon $H$ (Table 7 ), 80 decisions, $25 of API cost
Memory	at most 8 skeleton phases and 2 references; all records used for the calibrator
Cost	list prices per million tokens: $10 fresh input,$ 1 cached input, $12.5 cache write,$ 50 output (reasoning tokens are output). $J$ sums the cost of all attempts, failed ones included, and divides by the successes
Compute	one job per task and arm: 1 NVIDIA A40 (rendering, and serving Qwen3.8-27B), 8 CPU cores

Appendix

Table 5: Collection settings, shared by all tasks. The Qwen3.8-27B arms use the same prompts, tools and settings except the model.

Waypoint $k$	1	2	3	4	5	6	7	8	9
Stated $q_{k}$	0.90	0.70	0.90	0.90	0.90	0.80	0.65	0.65	0.70
Calibrated $\hat{p}_{k}$	0.88	0.54	0.85	0.84	0.84	0.67	0.40	0.42	0.53
Outcome	reached	reached	reached	reached	reached	stalled	deferred

Appendix

Table 6: The first plan of CloseFridge episode 5 (Fig. 1 ). The calibrator lowers the stated confidences; the commit rule executes waypoints 1–6 and defers 7–9, the first of which falls below

\tau=0.5

			GPT-6 Astra		Qwen3.8-27B		$/demo
Task	$H$	Success criterion	Base	CAPEX	Base	CAPEX	Base	CAPEX
Pick and place
CoffeeSetupMug	600	Mug within 4 cm (horizontal), 10 cm (vertical) of dispenser site; gripper $>$ 25 cm from object	3	6	0	0	8.14	5.80
PnPCounter $\to$ Cabinet	750	Object bounding box inside cabinet interior; gripper $>$ 25 cm from object	2	12	0	0	14.57	1.43
PnPCounter $\to$ Stove	600	Object touches pan, within 7 cm of its centre (horizontal); gripper $>$ 25 cm from object	0	20	0	3	—	0.41
PnPDrawer $\to$ Counter	750	Object touches a counter; gripper $>$ 25 cm from object	0	13	0	0	—	1.81

Appendix

Table 7: The 18 RoboCasa tasks and per-task collection results on the 20 recreated human-demonstration starts per task (successes out of 20 unless a denominator is shown, which counts starts with a valid episode).

H

: horizon in control steps at 20 Hz. Success criteria summarise RoboCasa’s _check_success , evaluated after every control step; joint positions are normalised to

[0,1]

over the joint range. $/demo: list-price API cost of all attempts divided by successes.

	Diffusion Policy (fine-tuned / scratch)	ACT
Inputs	3 RGB views at $256{\times}256$ ; end-effector position, quaternion and gripper joints
Actions	12-D at 20 Hz: end-effector deltas, gripper, base, control mode
Language	CLIP ViT-L/14 instruction embedding	none
Model	ResNet-18 encoders, transformer denoiser; DDPM, 100 steps; predicts 10 steps, executes 8	ResNet-18, transformer, CVAE (latent 32); chunk 20, temporal ensembling
Initialisation	RoboCasa365 multitask checkpoint / random	ImageNet backbone
Optimiser	AdamW, lr $10^{-4}$ , cosine, 500 warm-up steps, EMA	AdamW, lr $10^{-5}$ , constant

Appendix

Table 8: Policy recipes. DP follows RoboCasa365’s Diffusion Policy recipe; ACT uses LeRobot’s implementation. Both sources are trained identically.

	Base GPT-6 Astra	CAPEX
Prompt	interface only (the two-arm version of the simulation base prompt)	the simulation prompt plus YAM embodiment passages (grasp alignment, rim pinching)
Per decision	one waypoint	a plan of up to 24 waypoints, each with a stated confidence; committed up to the first later waypoint with stated confidence below 0.5
Perception	RGB only	RGB, and locate_pixel on calibrated depth
Memory	none	reliability skeleton and two of the run’s own successes with their start images

Appendix

Table 9: The two teachers on the YAM rig.

Task	Teacher	Success	95% CI	Min./success	Calls/success	$/success
Bread and corn	Base GPT-6 Astra	12/20	39–78%	3.33	25.7	1.10
	CAPEX	20/21	77–99%	1.49	7.5	0.43
Bowl handover	Base GPT-6 Astra	6/20	15–52%	7.58	61.5	2.73
	CAPEX	20/21	77–99%	2.26	15.8	0.88
Both	Base GPT-6 Astra	18/40		4.75	37.6	1.64
	CAPEX	40/42		1.87	11.7	0.65

Appendix

Table 10: Collection on the YAM rig. Success is the operator’s label, with a 95% Wilson interval. Time, model calls and API cost are totals over all attempts divided by the successes. Both teachers ended every attempt with done ; the operator rejected 22 of Base GPT-6 Astra’s 40 claims and 2 of CAPEX’s 42.

Panel of Figure 5 — Figure 5: Closed-loop rollouts on the Franka of Diffusion Policy trained only on 20 CAPEX demonstrations per task, six evenly spaced frames from the exterior camera (cropped to the workspace). Instructions: Pick Coke , “grab the coke can and lift it up”; Stack Cups , “stack the orange cup onto the green cup”; Stack Blocks , “stack the blue block on top of the green block”. Durations: Pick Coke 9 s, Stack Cups 34 s, Stack Blocks 38 s.

Panel of Figure 6 — Figure 6: Bimanual YAM, front camera, six evenly spaced frames per attempt. Top two rows: bread and corn to plate from the same start (trial 5). CAPEX places both items without moving the plate (32 s); Base GPT-6 Astra pushes the plate off its position, reports the task done, and the operator labels the attempt a failure (37 s). Bottom row: bowl handover, CAPEX (trial 0, 49 s): lift by the rim, hand-over in the air, set-down upright on the right.

Panel of Figure 7 — Figure 7: Demonstrations collected by CAPEX (top rows) and the human teleoperated demonstrations recreated from the same initial states (bottom rows), six evenly spaced frames each; for each task the CAPEX demonstration closest to the task’s median length is shown. These pairs are training examples of the downstream policy study (Sec. 4.4 ).

Explore similar work

ARC: A Reasoning Recipe for Robot Foundation Models
Oct 8, 2026Gokul Puthumanaillam, Tao Sun, Elie Aljalbout +5Robot Foundation Models Reasoning for VLA Models
University of Illinois Urbana-Champaign · NVIDIA · Stanford University +2
AXIS: A Growable Community-Driven Data Engine for Scalable Robot Manipulation
Jul 23, 2026Mengfei Zhao, Dihong Huang, Yikai Tang +12Robot Skill Learning Robotic Data Collection
1Axis Robotics · University of California, Berkeley · 3Georgia Institute of Technology +6
Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models
Jun 16, 2026Haoqi Yuan, Zhixuan Liang, Anzhe Chen +20Robot Skill Learning Robot Foundation Models
Qwen Team

Oct 8, 2026J/K move · Enter open · S save

ARC: A Reasoning Recipe for Robot Foundation Models

Gokul Puthumanaillam, Tao Sun, Elie Aljalbout, Moritz Reuss+4

University of Illinois Urbana-Champaign · NVIDIA · Stanford University+2

Robot Foundation Models Reasoning for VLA Models Zero-Shot Robotic Manipulation

Open paper AskPDF arXiv

The prevailing approach to improving robot foundation models (RFMs) relies on larger models, more robot demonstrations, and costly training at scale. We show that there exists an effective and efficient complementary approach: the right reasoning recipe can substantially improve the zero-shot task performance of existing state-of-the-art RFMs. We refer to this recipe as ARC. It consists of three key ingredients: a reasoning trace, a scalable automatic labeling pipeline, and a strategy for adapting pretrained RFMs to use these traces for control. First, we find that effective reasoning traces should be grounded in the robot's next action and explain its causal structure: why the action is appropriate and what effect it should produce. Second, we show that these traces can be generated automatically from existing demonstrations, enabling us to construct ARC-Trace-DROID from DROID without collecting new robot data. Third, we show how state-of-the-art VLAs such as

π_{0.5}

and WAMs such as Cosmos3-Nano-Policy can learn to use these traces for control, with fine-tuning and inference tailored to each model's architecture and capabilities. Using ARC, we obtain gains in zero-shot RFM performance that, to our knowledge, are unprecedented without additional robot demonstrations or foundation-scale training. The adapted models establish a new state of the art on RoboLab-120 and MolmoSpaces, with gains of up to 50 percentage points on RoboLab-Reasoning-50. On real robots, ARC improves

π_{0.5}

's task success by 82.2 percentage points. Project website: https://arc-robot-reasoning.github.io/

Abstract

Explore similar work

ARC: A Reasoning Recipe for Robot Foundation Models

AXIS: A Growable Community-Driven Data Engine for Scalable Robot Manipulation

Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models