DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning

Minho Park; Kinam Kim; Donghu Kim; Byungkun Lee; Dongyoon Hwang; Yongjae Shin; Junha Hyung; Hojoon Lee

cs.ROOct 23, 2026

DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning

Authors: Minho Park, Kinam Kim, Donghu Kim, Byungkun Lee, Dongyoon Hwang, Yongjae Shin, Junha Hyung, Hojoon Lee,

Organizations: KAIST · Holiday Robotics

Abstract

Behavior cloning for robotic manipulation is highly sensitive to the coverage of its training data, yet collecting diverse successful demonstrations remains costly. We study reinforcement learning (RL) not as a deployable controller, but as a data generator for downstream visuomotor imitation learning. Our key premise is that multiple successful solution modes cover a broader range of successful trajectories than repeatedly collecting a single narrow solution. We introduce DiscoDemo, which trains a privileged latent-conditioned RL generator with a parallel reverse curriculum and a success-gated diversity objective to discover diverse solutions while preserving task success. The trained generator produces multiple successful closed-loop behaviors, which are safety-filtered and exported as a static dataset for downstream imitation learning. Across four manipulation tasks, DiscoDemo produces broader successful-state coverage and improves downstream visuomotor policies in both simulation and the real world. Averaged across tasks, downstream success increases to 62% from 40% for the strongest baseline in simulation, and to 78% from 39% on the real robot, respectively. These results suggest that successful solution diversity improves imitation learning for a pretrained flow-matching VLA and remains complementary to corrective data collection. Project page: https://davian-robotics.github.io/DiscoDemo

Figures & tables

Figure 1: Solution discovery broadens coverage. Single-solution data provide narrow coverage, and augmentation mainly expands around existing trajectories. DiscoDemo , our method instead discovers distinct successful solutions that broaden state-space coverage.

Figure 2: Overview of DiscoDemo . (a) We train a privileged generator with a parallel reverse curriculum, sparse task reward, success-gated diversity, and safety shaping. (b) The trained generator produces successful and diverse demonstrations, filtered for physically undesirable interactions. (c) The retained observation–action data train a downstream visuomotor policy by imitation learning.

Figure 3: Downstream imitation performance in simulation and the real world. DiscoDemo achieves the highest success rate across all tasks with competitive completion times. Simulation: mean

\pm

s.d. over three fine-tuning seeds.

Figure 4: Qualitative Results of Generated Demonstrations. Each line traces the tool center point (TCP) of one successful demonstration from the same initial placement.

Figure 5: Dataset analysis. On average across tasks, DiscoDemo produces the most diverse and widest-coverage datasets, with high yield, short episodes, and low end-to-end GPU cost.

Figure 6: Effect of diversity strength $\alpha$ . Banana trajectory overlays and four-task averages diversity and SR. Diversity increases with $\alpha$ , while downstream success peaks at an intermediate value.

Figure 7: Generator training efficiency. P-RFCL and DiscoDemo complete training faster than PLD and ExpertGen. Times are estimated for the former and recorded for the latter ( Section F.4 ).

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Experimental Setup. The four tasks on the real FR3 (top) and in the simulated counterpart used for data generation (bottom), seen from the same exocentric camera.

Task	Horizon	Initial-state distribution $\rho_{0}$	Success condition
PnP-Banana	$30$ s	banana and bowl positions uniform in $x\in[0.41,0.64]$ m, $y\in[-0.15,0.15]$ m, at least $20$ cm apart; banana yaw uniform in $\pm 180^{\circ}$	banana center inside the bowl, in contact with the bowl, and released by the gripper
StackCube	$30$ s	red and blue cube positions in the same region, at least $15$ cm apart; red-cube yaw uniform in $\pm 45^{\circ}$	red cube on the blue cube: height offset within $1$ cm of $5$ cm, horizontal offset $\leq 2.5$ cm, speed $\leq 2$ cm/s
FMB-Round	$20$ s	peg grasped at the home configuration; fixture-board position uniform in $\pm 3$ cm and yaw in $\pm 15^{\circ}$	peg fully inserted (within $2$ mm), lateral error $\leq 5$ mm, axis misalignment $\leq 10^{\circ}$
FMB-SqCircle	$20$ s	as FMB-Round	as FMB-Round

Appendix

Table 2: Task specifications. Positions are in the robot base frame.

Hyperparameter	Value
Parallel environments	$2{,}048$
Environment steps	$50$ M
Actor / critic	$2$ residual blocks, width $128$ / $256$
Critic support	$101$ atoms on $[-5,5]$ , two critics
Optimizer	Adam, lr $3\!\times\!10^{-4}\to 1.5\!\times\!10^{-4}$ (cosine)
Batch size	$2{,}048$ ( $50\%$ demonstrations)

Appendix

Table 3: Generator hyperparameters, shared by P-RFCL and DiscoDemo and by all tasks.

	Original RFCL	Ours
Reverse-phase environments	$8$	$2{,}048$
Environment-to-demonstration	one pinned environment each	a group of $\approx M/N$ environments each
Advance criterion	$3$ consecutive successes	success rate $\geq 0.9$ over $10$ episodes
Curriculum stride	$4$	$16$
Forward phase	PLR over initial states	$\rho_{0}$ once half the demonstrations are solved
Success	at a single step	held for $1$ s

Appendix

Table 4: Our

\alpha=0

generator (P-RFCL in our experiments) compared with the original RFCL configuration.

Hyperparameter	Value
Initialization	DROID-pretrained $\pi_{0.5}$ (joint-position actions)
Trained parameters	vision encoder, language model, action expert
Frozen parameters	action/time projections, multi-modal projector, LM head
Precision	bfloat16
Optimizer	AdamW, $\beta=(0.9,0.95)$ , weight decay $0.01$ , grad. clip $1.0$
Learning rate	$1$ k warmup, cosine $2.5\!\times\!10^{-5}\to 2.5\!\times\!10^{-6}$

Appendix

Table 5: Downstream

\pi_{0.5}

fine-tuning recipe, identical for every data source.

Figure 10: Per-task dataset analysis. DiscoDemo attains the highest diversity and the highest or tied-highest coverage on three of the four tasks; on StackCube, PGDG exceeds it on both.

Figure 11: Coverage and diversity versus downstream success. Dataset coverage (top) tracks the success of the downstream policy more closely than fixed-scene diversity (bottom).

Task	Shaping	Successes	Passed	Pass rate (%)	Needed
PnP-Banana	✗	11,428	3,577	31.3	9,585
PnP-Banana	✓	10,730	5,358	49.9	6,008
FMB-SqCircle	✗	11,371	2,701	23.8	12,630
FMB-SqCircle	✓	11,632	5,395	46.4	6,468

Appendix

Table 6: Export-filter pass rates among successful rollouts. The last column is the number of successes needed for

3{,}000

filtered demonstrations.

Data source	SR (%) $\uparrow$	Time (s) $\downarrow$	Succ./h $\uparrow$	SR (%) $\uparrow$	Time (s) $\downarrow$	Succ./h $\uparrow$	SR (%) $\uparrow$	Time (s) $\downarrow$	Succ./h $\uparrow$	SR (%) $\uparrow$	Time (s) $\downarrow$	Succ./h $\uparrow$
	PnP-Banana			StackCube			FMB-Round			FMB-SqCircle
H20	0.0	30.0	0	0.0	30.0	0	0.0	20.0	0	0.0	20.0	0
MP ( 2026 )	20.0	17.1	26	0.0	30.0	0	40.0	14.6	81	5.0	14.1	9
MG ( 2023 )	0.0	30.0	0	50.0	20.9	71	25.0	14.6	48	20.0	12.8	39
PGDG ( 2026 )	0.0	30.0	0	50.0	19.1	73	30.0	13.6	60	20.0	14.1	38
PLD ( 2026 )	5.0	16.8	6	30.0	17.3	41	15.0	8.3	30	5.0	5.8	9

Appendix

Table 7: Real-world experiments.

	PnP-Banana			StackCube			FMB-Round			FMB-SqCircle
Data source	SR (%) $\uparrow$	Time (s) $\downarrow$	Succ./h $\uparrow$	SR (%) $\uparrow$	Time (s) $\downarrow$	Succ./h $\uparrow$	SR (%) $\uparrow$	Time (s) $\downarrow$	Succ./h $\uparrow$	SR (%) $\uparrow$	Time (s) $\downarrow$	Succ./h $\uparrow$
H20	1.3 $\pm$ 2.3	27.5 $\pm$ 4.3	2 $\pm$ 3	7.3 $\pm$ 5.0	21.2 $\pm$ 1.7	9 $\pm$ 6	0.0 $\pm$ 0.0	20.0	0 $\pm$ 0	0.0 $\pm$ 0.0	20.0	0 $\pm$ 0
MP ( 2026 )	1.3 $\pm$ 2.3	29.4 $\pm$ 1.1	2 $\pm$ 3	3.3 $\pm$ 3.1	22.6 $\pm$ 6.4	4 $\pm$ 4	6.0 $\pm$ 3.5	16.3 $\pm$ 1.6	11 $\pm$ 6	2.7 $\pm$ 3.1	16.7 $\pm$ 2.9	5 $\pm$ 6
MG ( 2023 )	41.3 $\pm$ 16.2	21.5 $\pm$ 2.5	58 $\pm$ 26	48.7 $\pm$ 11.0	18.5 $\pm$ 1.1	73 $\pm$ 20	31.3 $\pm$ 4.2	12.3 $\pm$ 0.1	64 $\pm$ 10	35.3 $\pm$ 13.0	9.6 $\pm$ 0.6	81 $\pm$ 37
PGDG ( 2026 )	27.3 $\pm$ 5.0	20.9 $\pm$ 1.0	36 $\pm$ 7	42.0 $\pm$ 9.2	17.2 $\pm$ 0.2	62 $\pm$ 16	15.3 $\pm$ 4.2	14.7 $\pm$ 1.2	29 $\pm$ 8	30.7 $\pm$ 9.9	14.0 $\pm$ 0.5	61 $\pm$ 22
PLD ( 2026 )	20.0 $\pm$ 3.5	23.4 $\pm$ 0.5	25 $\pm$ 4	37.3 $\pm$ 1.2	16.1 $\pm$ 0.8	54 $\pm$ 3	28.0 $\pm$ 3.5	6.3 $\pm$ 3.8	63 $\pm$ 12	8.0 $\pm$ 3.5	11.8 $\pm$ 0.9	15 $\pm$ 7

Appendix

Table 8: Simulation experiments.

Figure 12: The shape of the trade-off is common to the four tasks, but the

\alpha

at which generator success starts to give way is not.

Figure 13: Generated demonstrations on StackCube and FMB-Round, each line one successful demonstration from the same initial placement.

Figure 14: Data generators on PnP-Banana. Rollouts from the shared initial scene.

Figure 15: Data generators on StackCube. Rollouts from the shared initial scene.

Figure 16: Data generators on FMB-Round. Rollouts from the shared initial scene.

Figure 17: Data generators on FMB-SqCircle. Rollouts from the shared initial scene.

Figure 18: Downstream policies on PnP-Banana. Rollouts on the selected shared initial state.

Figure 19: Downstream policies on StackCube. Rollouts on the selected shared initial state.

Figure 20: Downstream policies on FMB-Round. Rollouts on the selected shared initial state.

Figure 21: Downstream policies on FMB-SqCircle. Rollouts on the selected shared initial state.

Figure 22: Real-robot rollouts on PnP-Banana. Rollouts on the scene used in the project-page video.

Figure 23: Real-robot rollouts on StackCube. Rollouts on the scene used in the project-page video.

Figure 24: Real-robot rollouts on FMB-Round. Rollouts on the scene used in the project-page video.

Figure 25: Real-robot rollouts on FMB-SqCircle. Rollouts on the scene used in the project-page video.

Robot Demonstration Generation

See all 13 papers

It's Not Just More Demos: Counterfactual Action Sensitivity Coverage for Data-Efficient Robust Robot Imitation
Jul 29, 2026Giovanni D'urso, Kaushik Roy, Nicholas Lawrance +1Visuomotor Policy Learning Imitation Learning
CSIRO Robotics, CSIRO, Australia
ReGIL: Retrieval-Guided Imitation Learning from a Single Demonstration
Jun 8, 2026Yuying Zhang, Francesco Verdoja, Wenyan Yang +1Robot Skill Learning Robotic RL
School of Electrical Engineering Aalto University, Finland
Learning Beyond What Humans Can Demonstrate
Sep 21, 2026Yuchen Song, Aditya Mittal, Unnat JainRobotic Data Collection Robotic Manipulation
University of California, Irvine

Jul 29, 2026J/K move · Enter open · S save

It's Not Just More Demos: Counterfactual Action Sensitivity Coverage for Data-Efficient Robust Robot Imitation

Giovanni D'urso, Kaushik Roy, Nicholas Lawrance, Brendan Tidd

CSIRO Robotics, CSIRO, Australia

Visuomotor Policy Learning Imitation Learning Robotic Manipulation Robot Imitation Learning

Open paper AskPDF arXiv

Visuomotor imitation learning has demonstrated success for manipulation tasks. However, the trained policies remain brittle to visual

nuisances', with even minor task-preserving variations such as lighting, distractions or changes in colour result in heavy degradation of the trained policy's performance. While increasing data diversity can improve robustness, it is unclear which additional demonstrations are informative for a particular trained policy. We propose Counterfactual Nuisance Behaviour Cloning (CFNBC), an offline data-selection framework for targeted robustness repair. Starting from a nominal policy trained on

clean' demonstrations, CFNBC generates paired clean and nuisance observations that preserve the expert action, then measures action drift: the change in the policy's predicted action under a nuisance that should not alter the desired behaviour. This provides a policy-specific sensitivity signal for selecting a compact, response-diverse repair set from a larger candidate pool, without requiring rollout success labels or online policy execution. We show in MuJoCo bimanual cube transfer and SimplerEnv cube stacking that action drift correlates with nuisance-induced failure, and that response-guided repair with only

20

30

selected candidates substantially outperforms matched-budget random selection while approaching the performance of much larger random repair budgets. These results support a data-centric view of robustness repair: the most useful data are not necessarily the most numerous, visually diverse, or obviously difficult, but the examples that cover fragile response modes of the current policy.

Abstract

Robot Demonstration Generation

Explore similar work

It's Not Just More Demos: Counterfactual Action Sensitivity Coverage for Data-Efficient Robust Robot Imitation

ReGIL: Retrieval-Guided Imitation Learning from a Single Demonstration

Learning Beyond What Humans Can Demonstrate