Robotic Data Inoculation

TL;DR

We introduce Data Inoculation (DI), a shared-control framework for collecting better robot demonstrations. A residual policy learns online to perturb human teleoperation commands, while the human continuously corrects these deviations. The resulting data contains recovery behaviors that “inoculate” the learner against its future mistakes—improving or matching downstream performance with both behavioral cloning and offline RL.

Can We Collect Better Data by Changing How the Robot Behaves?

Robot policies depend on the coverage of their demonstration data. Expert demonstrations show what success looks like, but often miss the near-failure states a learned policy will encounter. Small mistakes can compound, leaving the robot in unfamiliar states without a way to recover.

Online RL explores through autonomous trial and error. DAgger-style interventions wait for a policy to make a mistake before a human corrects it. What if we could proactively collect those corrections during demonstration?

The key idea. Let the human and a learned policy control the robot together. Expose the robot to controlled doses of policy-like mistakes, and record how the human recovers.

Our method

Data Inoculation

A human stays in the loop. A residual policy makes the data more useful.

1Perturb the Human’s Commands

At each step, the inoculative policy observes the robot and the human’s live command, then adds a bounded residual. The human reacts to the deviation through subsequent commands, pairing perturbations with corrective demonstrations.

Executed actionã_t=Human commanda_t+Learned perturbationα Δ_t

The residual is bounded so the executed action deviates from the human’s command by at most α in the L_∞ norm.

2Learn During Data Collection

An asynchronous online RL loop trains the residual policy on the growing dataset to maximize task reward. Because it shares the reward-maximizing objective used for downstream policy extraction, its deviations are expected to resemble the mistakes a learned policy would make.

Early critic errors lead to deviations that elicit recovery. As learning progresses, the residual policy can also assist the teleoperator. Perturbations are held fixed for 0.5 seconds, giving the human time to respond.

3Train on the Inoculated Dataset

Use the collected data to train an autonomous policy with behavioral cloning or offline RL. The dataset now includes both successful behavior and human corrections in states that clean teleoperation may rarely visit.

Why learn the perturbations? Random noise can spend the collection budget on irrelevant movements. DI adapts to the task, eliciting corrections where precision matters most.

Three Real-World Precision Tasks

We evaluate DI on a Franka robot arm with a Robotiq gripper and two wrist cameras. Each task requires careful manipulation, where small errors can determine success.

A sequence of robot poses picking up two cubes and balancing them on a yellow cube. — T-Cube
Balance two cubes diagonally on a third cube. A few millimeters of misalignment can topple the stack.

A sequence of robot poses inserting a USB-C cable through a narrow opening in a computer case panel. — Cable
Pick up a USB-C cable and insert it through a case opening barely larger than the connector.

A sequence of robot poses using an orange keycap puller to remove a key from a keyboard. — Keycap
Pick up a keycap puller, hook it onto the Q key, and use controlled motions to remove the keycap.

Better Data. More Successful Policies.

We compare DI with clean teleoperation and teleoperation with random perturbations. DI improves success rates on Cable and T-Cube, and performs similarly or slightly better on Keycap, under both behavioral cloning (BC) and offline RL.

Success rates for clean, random, and DI data across Cable, T-Cube, and Keycap, for both BC and RL. DI achieves the highest mean success in all six comparisons. Exact values follow in the expandable table. — Results averaged over 9 collection seeds for T-Cube and 5 each for Cable and Keycap, with 20 evaluation rollouts per seed. Error bars show standard errors.

35% → 67%Cable · offline RL

42% → 59%T-Cube · offline RL

81% → 84%Keycap · offline RL

Mean success rate: clean demonstrations → Data Inoculation.

Full success rates and standard errors

Success rate (%) ± standard error
Task	Policy	Clean	Random	DI
Cable	BC	37 ± 6	48 ± 9	66 ± 4
Cable	Offline RL	35 ± 7	56 ± 4	67 ± 6
T-Cube	BC	38 ± 4	40 ± 5	50 ± 3
T-Cube	Offline RL	42 ± 6	42 ± 3	59 ± 2
Keycap	BC	63 ± 9	69 ± 6	72 ± 5
Keycap	Offline RL	81 ± 6	67 ± 9	84 ± 6

During collection

From Perturbing to Assisting

On T-Cube, DI initially lowers collection success while the residual policy is untrained. As it learns, collection success rises and eventually exceeds clean and random collection. Meanwhile, the magnitude of the residual decreases.

T-Cube data collection curves: DI success increases during training and its residual action magnitude decreases. — The residual policy learns to make smaller, more useful deviations as data collection progresses.

Understanding the data

Corrections Where They Matter Most

We divide T-Cube trajectories into five phases and measure where human corrections occur. DI concentrates corrective demonstrations in place/release—the critical phase where imprecise motion can knock over the cubes. Random perturbations elicit many corrections during the less demanding retreat phase.

Correction distribution across approach, grasp, transport, place/release, and retreat, with robot images illustrating each phase. DI emphasizes place/release; random perturbations emphasize retreat. — Learning task-dependent perturbations helps collect recovery behavior at the moments that determine success.

Better demonstrations include how to recover. DI makes data collection a shared process: the policy exposes gaps, and the human supplies the corrections.

BibTeX

@misc{li2026robotic,
  title         = {Robotic Data Inoculation},
  author        = {Li, Qiyang and Zhou, Zhiyuan and Xu, Charles and Levine, Sergey},
  year          = {2026},
  eprint        = {2610.39054},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.39054}
}