Robots trained by imitation copy their teachers, flaws included. When a teleoperator hesitates, trembles, or takes a meandering route to a grasp, a behavior-cloning policy learns to do the same. This sets a ceiling: the policy is at best as good as the data. This note proposes a simple intervention that sits between data collection and policy training. Before training, each human demonstration is split into free-space "transit" segments and contact-rich "interaction" segments. Every transit segment is then rewritten by a batched GPU trajectory optimizer. The rewrite keeps the human's choices: where they grasped, where they placed, and which side of each obstacle they went around. It discards their execution noise, meaning jitter, pauses, and slow detours. The repaired episode is replayed in simulation to regenerate observations and to confirm it still reaches the same post-segment state, and only then is a standard visuomotor policy trained on it.
The bet is that, on mixed-skill human data, much of the gap between a mediocre and a proficient demonstrator lies in low-level execution rather than strategy. If so, it can be closed with collision-aware motion planning and no reward function. Robomimic's paired Proficient-Human (one expert) and Multi-Human (six operators of varying skill) splits make this directly testable. The prediction is that a policy trained on repaired Multi-Human data should recover much of the success-rate and completion-time gap to one trained on Proficient-Human data. It should also beat both naive smoothing and pure planner-generated data.
Core Claim
Standard imitation assumes the demonstrator is optimal. The methods that relax this assumption mostly reweight or discard whole trajectories. Advantage weighting needs rewards (Exponentially Weighted Imitation Learning for Batched Historical Data). Confidence or expertise estimators score trajectories without rewards (Confidence-Aware Imitation Learning from Demonstrations with Varying Optimality, Imitation Learning by Estimating Expertise of Demonstrators). The research file notes the cost of this approach: a demonstration with sound intent but sloppy execution is down-weighted wholesale, and its intent is lost. The claim here is that these two components can be separated surgically. The demonstration supplies segment endpoints and a route through the obstacles. A GPU optimizer supplies the execution. A single knob trades fidelity to the human route against motion quality.
For a transit segment with human path $\xi_h$, the repaired path $\xi$ solves
$$\min_{\xi}\; c_{\text{time}}(\xi) + c_{\text{jerk}}(\xi) + \lambda\, D_{\text{spec}}(\xi, \xi_h) \quad \text{s.t. } \xi \text{ collision-free, within joint limits, } \xi(0)=\xi_h(0),\ \xi(1)=\xi_h(1).$$
Here $c_{\text{time}}$ penalizes duration and $c_{\text{jerk}}$ penalizes the third derivative of joint motion. $D_{\text{spec}}$ is the mean-squared error between low-frequency spectra of the two paths after time-normalization. It is the spectral shape metric from Stochastic Trajectory Optimization for Demonstration Imitation, which that paper shows upper-bounds dynamic time warping at linear cost. The weight $\lambda$ sets how closely the human route is kept. At $\lambda = 0$ the method reduces to replacing the human with a planner. At very large $\lambda$ it reduces to collision-aware smoothing. The central empirical question is whether an intermediate $\lambda$ beats both ends.
Research Program
The pipeline has four stages, all in Robomimic simulation, where scene geometry is known.
The first stage is segmentation. Contact flags and gripper state from the simulator mark interaction phases. These are any phase where the robot or a held object touches anything other than the grasped object. Everything else is transit. Transit includes carrying a rigidly grasped object, which is handled as an attached collision body.
The second stage is batched repair. For each transit segment, the optimizer from CuRobo: Parallelized Collision-Free Robot Motion Generation is extended with the $D_{\text{spec}}$ term. Many seeds are run in parallel, all initialized from perturbations of $\xi_h$, across a sweep of $\lambda$. The solution that best fits the objective is kept. The GPU matters here for a specific reason. The method needs dozens of seeds times several $\lambda$ values times thousands of segments, plus optional augmentation with several distinct repaired variants per segment. A single-instance CPU optimizer like MSTOMP is not built for this. cuTAMP reports nearly constant runtime from 1 to 1024 particles (Differentiable GPU-Parallelized Task and Motion Planning), which is the batching property being relied on. Interaction segments are only lightly processed: pauses longer than a threshold are removed, and nothing else changes.
The third stage is replay verification. The stitched action sequence is executed in the simulator. A repair is accepted only if the state at each segment boundary matches the original demonstration within tolerance. Rejected repairs fall back to the raw segment. This check uses the demonstration itself as the reference, not a task reward, so no reward engineering is introduced. The acceptance rate is itself a reported result.
The fourth stage is policy training. A Diffusion Policy (Diffusion Policy: Visuomotor Policy Learning via Action Diffusion), meaning a generator that produces action sequences by iteratively denoising random noise, is trained on the repaired data. BC-RNN, the recurrent behavior-cloning model favored for human data, is also trained on it. Training recipes are unchanged, so any gain is attributable to the data.
Prior work plays three roles here. MSTOMP supplies the shape metric and the motivating observation that human demos have good shape but poor dynamics. It stays a per-instance optimizer with no learned policy, and this proposal amortizes its idea into training data. DiffusionSeeder: Seeding Motion Optimization with Diffusion for Rapid Motion Planning is the reverse coupling: a diffusion model trained on 15 million synthetic planner trajectories seeds cuRobo. It never touches human data, and its $\lambda = 0$-like data source is one of the ablations here. DiffTORI: Differentiable Trajectory Optimization for Deep Reinforcement and Imitation Learning embeds a local optimizer inside the policy and fits its cost to demonstrations. That fit assumes the demonstrations are optimal, so it inherits their flaws. Here the optimizer acts on the data and is explicitly allowed to disagree with the human about execution.
Baselines
The naive reference point is Diffusion Policy and BC-RNN trained on raw Multi-Human data with the standard recipe. Proficient-Human training serves as a reference ceiling: it shows what one skilled operator achieves, and the proposal aims to approach it using worse operators.
The strongest existing number for human-data imitation in the file comes from What Matters in Learning from Offline Human Demonstrations for Robot Manipulation. It reports BC-RNN success rates of 84% on Square and 71.3% on Transport. The file does not specify the split for these numbers, so all runs here are re-trained with that paper's settings on both splits rather than relying on the quoted figure. No reported result was found on Lacuna for Diffusion Policy on Robomimic Multi-Human; the file gives only its 95% on Push-T. For noisy crowdsourced data specifically, IRIS reports 28.3% on RoboTurk Cans with multi-human demos, against 0% for its baselines, and serves as the hierarchical-stitching competitor where comparable.
Four further baselines come from the closest competing approaches. Confidence-weighted imitation (Confidence-Aware Imitation Learning from Demonstrations with Varying Optimality) is the strongest reward-free method for mixed-quality data, because it addresses exactly the varying-optimality problem by reweighting rather than repairing. Classical smoothing applies a Butterworth low-pass filter plus pause removal to transit segments with no collision check. A planner-replacement ablation sets $\lambda = 0$, so the human contributes only segment endpoints. A collision-aware smoothing ablation uses large $\lambda$. These last two isolate the key ingredient, the intent-preserving anchor.
All methods share the same policy architecture, the same number of gradient steps, the same evaluation seeds, and the same hyperparameter search budget. That budget is a fixed number of training runs per method, with $\lambda$ counted as a hyperparameter of the proposal. Success means beating raw Multi-Human training and confidence weighting by a margin exceeding seed variance on Square, Transport, and ToolHang. It also means beating both $\lambda$ extremes, which is what shows the anchor matters.
Why not just…?
Why not just low-pass filter the demonstrations? Filtering removes tremor but not detours or slow velocity profiles. MSTOMP reports that blind smoothing cuts corners across obstacles and can break kinematic constraints. Replay verification measures this directly: if a filter's acceptance rate and downstream success match the repair's, the planner is unnecessary, and the proposal will report that.
Why not just train on planner-generated demonstrations and ignore the humans? This is the most serious threat. DiffusionSeeder shows planner data at scale works for free-space motion. The $\lambda = 0$ ablation tests it directly. It keeps human grasp and place poses but discards the human route, and it can be run with as many synthetic variants as compute allows. If it matches intermediate $\lambda$, then human intent is irrelevant in transit. The claim would then shrink to "use humans only for interaction phases," which is still a useful finding but a smaller one.
Why not just use offline RL or value steering to exceed the demonstrator? These methods need reward signals and are fragile on human data. What Matters reports BCQ and CQL near 0–10% on human Square and Transport. IRIS reports batch RL at 18.0% on fumbled Lift demos. Test-time steering such as Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning and Steering Your Generalists is complementary rather than competing, since it reranks proposals from whatever policy it is given. One extra run applies V-GPS-style reranking on top of raw-trained and repair-trained policies, to test whether the two gains add.
Why not just train a bigger or more expressive policy on raw data? Diffusion Policy's authors note that it remains bounded by demonstration quality. Including Diffusion Policy as the raw-data baseline settles whether expressiveness alone closes the gap between the two human splits.
Proposed Experiments
The main experiment trains on Robomimic Lift, Can, Square, Transport, and ToolHang. Each task is trained on raw Proficient-Human, raw Multi-Human, and repaired Multi-Human data, with low-dimensional and image observations. Metrics are success rate over at least three training seeds, episode length, mean jerk, and minimum obstacle clearance during transit.
The first decisive check is the gap-closing ratio: the fraction of the Proficient-Human minus Multi-Human success gap recovered by repair. A ratio near zero falsifies the claim that execution noise is the ceiling. The second is the $\lambda$ sweep. Success should peak at an intermediate value, with the peak higher on Multi-Human than on Proficient-Human, because the expert has less noise to remove. A monotone curve favoring $\lambda = 0$ would redirect the work toward planner data. The third is a per-operator analysis on Multi-Human. If repair helps most on demos from the lowest-skill operators, as measured by original episode length, that supports the execution-noise hypothesis over generic augmentation.
Two diagnostics accompany these. One is the replay-acceptance rate per task, reported because ToolHang's tight clearances may make repair rarely accepted. The other is wall-clock repair time per demonstration on one GPU, against a CPU STOMP-style implementation, to show whether batching is actually needed at this scale.
References
- Exponentially Weighted Imitation Learning for Batched Historical Data
- Confidence-Aware Imitation Learning from Demonstrations with Varying Optimality
- Imitation Learning by Estimating Expertise of Demonstrators
- Stochastic Trajectory Optimization for Robotic Skill Acquisition From a Suboptimal Demonstration
- CuRobo: Parallelized Collision-Free Robot Motion Generation
- Differentiable GPU-Parallelized Task and Motion Planning
- Diffusion Policy: Visuomotor Policy Learning via Action Diffusion
- DiffusionSeeder: Seeding Motion Optimization with Diffusion for Rapid Motion Planning
- DiffTORI: Differentiable Trajectory Optimization for Deep Reinforcement and Imitation Learning
- What Matters in Learning from Offline Human Demonstrations for Robot Manipulation
- IRIS: Implicit Reinforcement without Interaction at Scale for Learning Control from Offline Robot Manipulation Data
- Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning
- Steering Your Generalists: Improving Robotic Foundation Models via Value Guidance