I3L (Iterative Interactive Imitation Learning)

I3L (Iterative Interactive Imitation Learning) is a robot learning system from researchers at Stanford University and Lambda Inc. that improves whole-body mobile manipulation policies by letting a human operator take over a running robot, correct it, and hand control back, then retraining the policy on those corrections over several rounds. It pairs a low-cost, open-source leader-arm interface called JoyLo+ with an interactive imitation learning loop based on HG-DAgger. The paper, "I3L: Iterative Interactive Imitation Learning for Whole-Body Mobile Manipulation," was submitted to arXiv on 13 October 2026 (arXiv:2610.17748), with Wensi Ai, Christopher Luey and Karthik Dharmarajan as equal-contribution first authors and Fei-Fei Li and Ruohan Zhang as equal advisors.^[1][2] Ai announced the work on X on 29 October 2026.^[5]

The authors evaluated I3L on eight simulated and real-world tasks across three robot embodiments, including four long-horizon household tasks on a real Galaxea R1 Pro wheeled dual-arm robot. In their experiments, four rounds of correction raised average autonomous success from 25.6% for the base policies to 93.9%, with at least 90% on every task, and beat an equal budget of extra demonstrations by 35.6 percentage points on average.^[1] These figures are self-reported by the authors.

Background

Policies trained by behavioral cloning on a fixed set of teleoperated demonstrations tend to fail once small errors push the robot into states the demonstrations never covered, a form of distribution shift usually described as compounding error or covariate shift.^[1][6] DAgger (Ross, Gordon and Bagnell, AISTATS 2011) addressed this by iteratively collecting expert action labels on the states the learner itself visits and aggregating them into the training set.^[6] HG-DAgger (Kelly et al., ICRA 2019) adapted DAgger for human experts, who find it hard to label actions without being in control of the system: in HG-DAgger the human decides when to intervene and provides corrections while in control.^[1][7] Later human-in-the-loop work applied similar ideas to robot manipulation, including IWR (Mandlekar et al. 2020), which balances intervention and non-intervention samples, and Sirius (Liu et al., RSS 2023), which reweights deployment data using an estimate of human trust.^[1][8][9]

The I3L paper argues that the problem is especially acute for whole-body mobile manipulation, which it describes as coupling "over twenty actuated degrees of freedom with an omnidirectional base," and that systematic evidence for corrective learning on long-horizon tasks that need coordinated base, torso and arm motion had been limited.^[1] The project builds on two earlier teleoperation interfaces: GELLO, a low-cost kinematic leader-arm teleoperation framework, and JoyLo, the whole-body teleoperation interface introduced with the BEHAVIOR Robot Suite (BRS), a framework for household manipulation built on a bimanual wheeled robot with a 4-DoF torso.^[1][10][11] Several I3L authors, including Yunfan Jiang, Ruohan Zhang, Jiajun Wu and Fei-Fei Li, were also authors of BRS.^[1][10]

JoyLo+ interface

JoyLo+ is an actuated leader-arm device whose motorized arms follow the robot's learned policy during autonomous execution; the operator takes over by grasping it. The paper describes two versions: an R1 Pro version with two seven-joint leaders for bimanual whole-body mobile manipulation, and a Franka version with one seven-joint leader for single-arm tabletop work.^[1] Its main design features are:

• Policy-tracking leaders. During autonomous execution, motorized leader arms follow the policy's arm joint targets, so the device is already aligned with the robot when the operator steps in. During a correction, the leaders switch to compliant current control with gravity compensation for hand guidance.^[1][2]
• Touch-triggered takeover. Copper electrodes on the grips, an Adafruit MPR121 capacitive-sensing board and an ESP32-S3 controller report touch state, so grasping the device hands control to the operator without a separate button press. Separate touch and release thresholds plus a short release delay keep brief losses of contact from returning control to the policy.^[1]
• Whole-body takeover. On the R1 Pro, touching a grip or using a Joy-Con input transfers command of both arms, the torso, the mobile base and the grippers together. The leaders drive the arms and the Joy-Cons drive the torso, base and grippers. Releasing the device returns control to the policy, which replans from the corrected state.^[1][2]

The authors report the following component costs (robot, tax, shipping, cabling, fasteners, bearings and mounting hardware excluded):^[1][2]

The project website publishes the bill of materials, downloadable 3D-printable part files and 3D models for both versions.^[2]

Learning procedure

I3L starts from a demonstration dataset and runs repeated rounds of deployment and retraining. In each round the current policy runs on the robot, the operator intervenes through JoyLo+ when it approaches a failure, and every timestep is recorded with an intervention label: original demonstration, active intervention, a 15-frame pre-intervention window, or autonomous execution.^[1] Under HG-DAgger filtering, which the authors adopted after it performed best in their simulation comparison, only the original demonstrations and the active-intervention frames from all rounds are kept for training; pre-intervention and autonomous frames are dropped.^[1][2]

After each round the policy is retrained on the aggregated data from the same starting point rather than fine-tuned from the previous round: the same pretrained checkpoint for the vision-language-action model, and the same initial weights for diffusion policies. The updated policy is then deployed to collect the next round.^[1] The authors report that as early failures become rarer, corrections shift toward later stages of each task such as transport and placement, which they describe as an "emergent failure-mode curriculum."^[2]

Policies and training

Sources: paper Sections IV-A and Appendix E;^[1] the Franka Panda model is named in the project's code repository.^[3]

The diffusion policies are trained for 200,000 steps and the π0.5 policies for 40,000 steps before evaluation, and the last checkpoint is evaluated.^[1] The paper describes its π0.5 configuration^[12] as a PaliGemma backbone (gemma_2b language-model configuration) plus a gemma_300m action-expert configuration, about 3.35 billion parameters for the full model including the vision encoder, with each task starting from 100 base demonstrations; a 40,000-step run takes about seven hours on a single node of eight NVIDIA H100 GPUs. Real-robot policies were deployed on machines with NVIDIA RTX 4090 GPUs.^[1]

Tasks and embodiments

The evaluation covers eight tasks with 29 annotated subtasks in total. The two simulation tasks run in the OmniGibson simulator from the BEHAVIOR project, where scripted oracles using cuRobo motion planning generate the demonstrations and corrective motions while a human decides when to intervene and when to resume. The four R1 Pro tasks are real-world tasks from the BEHAVIOR benchmark.^[1][13]

According to the paper, the real R1 Pro tasks average about 83 seconds per execution, with breakfast preparation and book storage close to two minutes each. Real-world conditions are evaluated with 25 rollouts each; simulation results average five evaluation seeds of 50 rollouts.^[1]

Results

Corrections versus extra demonstrations

The main comparison gives each base policy the same additional budget of 100 trajectories, either as new expert demonstrations or as corrective trajectories collected over four HG-DAgger rounds of 25 episodes.^[1]

Source: paper Table II.^[1]

Across the eight tasks, extra demonstrations improved base success by 32.7 percentage points on average, while four-round HG-DAgger improved it by 68.3 points, a 35.6-point advantage. On the four real-world BEHAVIOR tasks alone, the paper reports an average gain of 81 points over the base policies and a 30-point advantage over extra demonstrations.^[1] In a separate simulation comparison with the same budget and HG-DAgger filtering, collecting corrections in four rounds of 25 (retraining between rounds) beat a single batch of 100 collected against the fixed base policy by 3.6 points on Strawberries and 10.0 points on Open Fridge.^[1]

Intervention burden

From round 1 to round 4, averaged over all eight tasks, the mean number of takeovers per episode fell by 63.4%, the share of frames under human control fell by 69.5%, and total takeover time per episode fell by 72.0%; all three measures declined in every round.^[1] Remaining corrections clustered at task-specific bottlenecks. On Boxing Books, for example, intervention around picking up books dropped substantially by round 4 while a peak remained around picking up the basket, which the authors attribute possibly to the difficulty of coordinating both grippers with the basket handles alongside base, torso and arm motion.^[1]

Whole-body corrections

On the four R1 Pro tasks, base movements accounted for an average of 34.2% of body-part activations during corrections, and torso movements for an average of 8.2% in the three tasks where they occurred. The authors cite this as evidence that operators corrected more than arm motion and that a whole-body takeover interface is useful.^[1]

Correction-learning strategies

In simulation, the authors compared four ways of learning from corrections under the same trajectory budgets: a residual policy that learns additive corrections on top of a frozen base policy, behavior cloning on all collected data, Sirius-style reweighting, and HG-DAgger filtering.^[1]

Mean success (%) over five evaluation seeds; source: paper Table III, which also lists rounds 1 to 3 and standard deviations.^[1]

Four-round HG-DAgger had the highest final mean on both tasks. The paper notes that more rounds alone did not guarantee a better policy: behavior cloning on all data came close on Open Fridge but did poorly on Strawberries, which the authors link to learning suboptimal behavior from on-policy data, and Sirius fluctuated across rounds on Strawberries.^[1][2]

Operator study

Three teleoperators with different levels of experience each ran three correction rounds of 25 episodes on Franka Mug Hanging (225 episodes in total), each refining their own policy. Cumulative intervention share fell for all three, from 16.1% to 11.8%, 34.6% to 19.9%, and 16.4% to 14.7%, and by round 3 takeover-free episodes made up 40-55% of each operator's cumulative collection.^[1]

Limitations

The authors list several limitations. Because multi-round whole-body experiments are expensive, the algorithm comparisons were run only in simulation, and the study covers eight tasks and three operators; broader task and operator coverage is left to future work. Each round retrains from a common initialization, and the authors suggest incremental updates could cut retraining cost. The intervention metrics measure active takeover time only, although operators also supervise the robot throughout execution; the paper says future work will quantify total operator time.^[1]

Code, models and data

The code is published at github.com/i-3l/I3L under the Apache License 2.0. According to its README, the repository includes the JoyLo+ and GELLO bilateral control code, Joy-Con input, simulated A1, R1 Pro and Franka environments in OmniGibson, real Franka Panda and Galaxea R1 Pro interfaces, HDF5 recording with intervention labels, LeRobot conversion tools, and sampling strategies for HG-DAgger, Sirius and co-training. It depends on a dedicated branch of the StanfordVL BEHAVIOR-1K repository and on the il_lib training library.^[3] As of 1 November 2026 the GitHub repository had 7 stars, and the I3L organization on Hugging Face hosted 24 model repositories and 36 dataset repositories.^[3][4]

Authors

The arXiv paper lists only Stanford University and Lambda Inc. as affiliations; the project website additionally lists Zhang with Northwestern University.^[1][2] Zhang's personal website describes him as an incoming assistant professor of computer science at Northwestern, joining in fall 2026.^[14] The work was posted as an arXiv preprint in the Robotics (cs.RO) and Artificial Intelligence (cs.AI) categories.^[1]

Continue reading

• IImitation LearningImitation learning is a family of methods for learning sequential behavior from demonstrations.Mentioned 3 times above
• HHuman-in-the-loopHuman-in-the-loop (HITL) describes any arrangement in which a person is a required participant in an automated system's operating cycle rather than a bystander to it.Listed in See also
• RRobot learningRobot learning studies how robots acquire or improve behavior from data and experience.Listed in See also
• WWhole-body controlWhole-body control (WBC) is a class of robotics control techniques that coordinates all of a robot's degrees of freedom at once to achieve multiple prioritized tasks, such as balancing, reaching, and…Listed in See also
• EEmbodied AIEmbodied AI is an area of artificial intelligence concerned with agents whose perception, action, learning, and reasoning are coupled through interaction with an environment.Listed in See also
• Stanford UniversityStanford University, legally Leland Stanford Junior University, is a private, nonprofit research university in Stanford, California.Mentioned 11 times above

More to read

• Fei-Fei Li
• Galaxea R1 Pro
• Teleoperation
• π0.5
• Vision-language-action model
• Diffusion policy

Topics:AI ResearchEmbodied AIMachine LearningRobotics

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35Ai, W., Luey, C., Dharmarajan, K., Han, E., Jiang, Y., Ren, S., Xie, J., Wu, J., Fei-Fei, L., Zhang, R. "I3L: Iterative Interactive Imitation Learning for Whole-Body Mobile Manipulation." arXiv:2610.17748, submitted 13 October 2026. arxiv.org/...2610.17748
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9"I3L: Iterative Interactive Imitation Learning for Whole-Body Mobile Manipulation." Project website. i-3l.github.io
3. ^1 ^2 ^3"i-3l/I3L." GitHub repository. github.com/...I3L
4. ^"I3L." Hugging Face organization. huggingface.co/I3L
5. ^Wensi Ai (@wensi_ai). "Introducing I3L: Iterative Interactive Imitation Learning for Whole-Body Mobile Manipulation." X, 29 October 2026. x.com/...2115885219815493909
6. ^1 ^2Ross, S., Gordon, G. J., Bagnell, J. A. "A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning." AISTATS 2011. arXiv:1011.0686. arxiv.org/...1011.0686
7. ^Kelly, M., Sidrane, C., Driggs-Campbell, K., Kochenderfer, M. J. "HG-DAgger: Interactive Imitation Learning with Human Experts." ICRA 2019. arXiv:1810.02890. arxiv.org/...1810.02890
8. ^Mandlekar, A., Xu, D., Martín-Martín, R., Zhu, Y., Fei-Fei, L., Savarese, S. "Human-in-the-Loop Imitation Learning using Remote Teleoperation." arXiv:2012.06733, 2020. arxiv.org/...2012.06733
9. ^Liu, H., Nasiriany, S., Zhang, L., Bao, Z., Zhu, Y. "Robot Learning on the Job: Human-in-the-Loop Autonomy and Learning During Deployment." RSS 2023. arXiv:2211.08416. arxiv.org/...2211.08416
10. ^1 ^2Jiang, Y., Zhang, R., Wong, J., Wang, C., Ze, Y., Yin, H., et al. "BEHAVIOR Robot Suite: Streamlining Real-World Whole-Body Manipulation for Everyday Household Activities." CoRL 2025. arXiv:2503.05652. arxiv.org/...2503.05652
11. ^Wu, P., Shentu, Y., Yi, Z., Lin, X., Abbeel, P. "GELLO: A General, Low-Cost, and Intuitive Teleoperation Framework for Robot Manipulators." IROS 2024. arXiv:2309.13037. arxiv.org/...2309.13037
12. ^Physical Intelligence, Black, K., Brown, N., et al. "π0.5: a Vision-Language-Action Model with Open-World Generalization." arXiv:2504.16054, 2025. arxiv.org/...2504.16054
13. ^Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martín-Martín, R., et al. "BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation." arXiv:2403.09227, 2024. arxiv.org/...2403.09227
14. ^Ruohan Zhang. Personal website. ruohanzhang.com

Show all 14 references