When a person controls a robot through a noisy channel, such as a brain-computer interface, a myoelectric armband or a laggy joystick, an assisting "copilot" can help by correcting bad commands. Existing diffusion copilots decide when to step in by asking whether the human's command looks bad for the task. Safety-critical systems need something different: a guarantee that a separate, explicitly measured cost (collisions, hazard entries, excessive force) stays under a fixed budget, and otherwise no interference with the user. We propose a copilot that stays silent while the human's command keeps the episode within its remaining safety budget. When a command would exceed that budget, it substitutes the nearest safe action drawn from a multi-modal model of expert behavior, and it commits to one way around an obstacle rather than flipping between alternatives. Here a diffusion model means a generator that produces actions by gradually denoising random noise, and multi-modal means it can represent several distinct good actions, such as "go left" and "go right," without averaging them into a crash.
Evaluation is the second half of the contribution. The research file contains strong evidence that offline scores do not predict online success in feedback loops. Less is more found no correlation between offline reconstruction quality and closed-loop cursor performance. We therefore treat closed-loop validation as a measured object. Every copilot is scored both by replaying logged pilot commands (open loop) and with pilots who react to the copilot's corrections (closed loop), including pilots who are still learning, and we report where the two rankings disagree. In this joint system the learner doing the exploring is the human, so "safe exploration" means keeping the budget while the operator is still learning to use the interface.
Core Claim
Triggering intervention on remaining cost budget rather than on task value, and sampling the replacement from a mode-committed diffusion copilot, gives two things. It lowers cumulative constraint violations in closed loop, including while the pilot is learning. It also intervenes less often and less jerkily than three alternatives at matched violation rates: reward-based intervention, cost-penalized reward intervention, and an external nearest-safe-action filter. We further claim that open-loop replay systematically overstates the benefit of always-on assistance and understates mode chattering, so copilot comparisons must be made closed loop.
Relation to Prior Work
The closest work is Shared Autonomy with IDA. IDA trains a goal-agnostic diffusion action refiner and intervenes only when the human's action is worse than the copilot's under a simulator-trained expert Q-function for every possible goal. It reports Lunar Lander success rising from 14% (pilot alone) to 91.7%, versus 68.2% for always-on assistance. IDA supplies our intervention-only-when-necessary structure, but it handles safety only through task return. We replace its reward-Q test with an explicit constrained criterion, in the sense of a constrained Markov decision process (CMDP): maximize reward subject to expected cumulative cost staying below a budget $\epsilon$. To the Noise and Back supplies the projection mechanism. It partially noises the human command and denoises it toward the expert manifold, using a globally fixed blending ratio, and it cut simulated laggy-pilot crash rates from 77% to under 10%. We keep partial denoising but make its activation depend on cost and budget rather than being always on.
A second group motivates the safety half but also constrains it. FISOR, DRCORL, ALGD and PACT all put safety into a diffusion policy, using reachability guidance, score regularization, augmented-Lagrangian guidance and post-training distillation respectively. All four act autonomously and have no human whose intent should be preserved. We reuse two pieces from this group: DRCORL's conservative upper-confidence cost-critic ensemble, and PACT's evidence that bolted-on test-time guardrails produce jerky trajectories. The Safety Paradox constrains the design. A cost critic that rarely sees violations misjudges the boundary, and an offline-trained critic is exactly such a critic. We therefore report violations separately for states inside and outside the training data's support and do not claim a guarantee beyond critic accuracy.
A third group motivates the closed-loop lens. NDT2 reports jerky "pulsar" movements when expressive decoders run online. Less is more shows that collapsing continuous velocity output to nine discrete options raised human closed-loop target hits to 93%, compared with 26–56% for continuous decoders. FALCON evaluates only on static splits. Read together, these results suggest that much of the closed-loop benefit comes from committing to a few stable options. Our mode-commitment component is a learned, continuous version of that idea, and we test it against DDS-style discretization directly.
Research Program
The copilot has three learned parts, all trained offline on DSRL trajectories for the navigation tasks and on simulator rollouts for Lunar Lander. The first is a state-conditioned diffusion model of safe expert actions. The second is an ensemble of cost critics, each estimating the expected future cost of taking action $a$ in state $s$; we take their upper confidence value $\bar Q_c(s,a)$. The third is a running record of cost already spent in the episode. At each step the human command $a_h$ passes through untouched if
$$\bar Q_c(s_t, a_h) \le b_t, \qquad b_t = \epsilon - \sum_{\tau < t} \gamma^{\tau - t}\, c(s_\tau, a_\tau),$$
where $b_t$ is the remaining budget, $\epsilon$ the episode budget, $c$ the per-step cost and $\gamma$ the discount. If the test fails, the copilot partially noises $a_h$, denoises $K$ candidates, and discards those violating the same inequality. Among the survivors it picks the one minimizing distance to $a_h$ plus a penalty for leaving the mode it committed to on the previous intervention.
Mode commitment is implemented by warm-starting denoising from the previous step's partially denoised action chunk, as in Falcon. This also brings sampling cost close to real time. If no candidate is feasible, the copilot executes the candidate with the lowest cost estimate and logs the event as a budget breach risk.
Closed-loop pilots come at three levels of realism. The first is scripted surrogate pilots: noisy, laggy and goal-switching, following the protocols in the two shared-autonomy papers. The second is a simulated decoder pilot. An intended velocity is encoded into synthetic population activity, decoded by a Kalman-style decoder, and the simulated user corrects on the basis of the observed cursor. This reproduces the feedback-driven input shift that defines the offline-to-online gap. The third is learning pilots: a PPO agent acting through the noisy decoder, trained while assisted. This tests safe exploration of the joint system and asks whether assistance blocks the pilot from improving. A small human study on Lunar Lander and Safety-Gymnasium PointGoal with game controllers, matching IDA's eight-participant scale, closes the program. If the jaBCI platform becomes available to us, it would be the natural next step; we do not assume it.
Baselines
The naive reference is the pilot acting alone. It sets the floor for violations and the ceiling for autonomy, since there are no interventions. On Lunar Lander the reported number is 14% success for human pilots alone in IDA.
The strongest existing method is IDA itself, run with its own reward-Q intervention rule. We use the same diffusion refiner weights as our copilot, so the only difference is the trigger and selection rule. Its reported closed-loop figure is 91.7% Lunar Lander success with eight human participants. On the Safety-Gymnasium navigation tasks we found no reported shared-autonomy result on Lacuna, so we will establish IDA's numbers there ourselves. Beating IDA means a violation rate at or below $\epsilon$ in at least as many tasks, with no more than a few points of success lost and equal or fewer interventions per episode.
The second competitor is partial-diffusion assistance To the Noise and Back, with its blending ratio swept over a grid. We report its best setting by success subject to the budget, so it receives a tuning budget equal to ours.
The third competitor is an external safety filter. It replaces any command failing the budget test with the closest action, found by sampling, that passes the same critic test. The filter has no diffusion model and no mode commitment, and it uses our exact critic. The fourth competitor is DDS-style discretization of pilot commands to nine directions before the same filter is applied, following Less is more.
Ablations remove one ingredient at a time. The first replaces the remaining-budget test with a fixed per-step threshold. The second removes mode commitment. The third substitutes a unimodal Gaussian copilot of equal parameter count. As a calibration reference for the critic, we run the offline-safe autonomous policies FISOR and DRCORL, which satisfy constraints on DSRL; DRCORL reports a normalized return of 0.91 on CarGoal1 against 0.38 for the next safe baseline. These run without any human, which tells us how much task competence the human-preserving constraint costs.
All methods share the training data, critic ensemble, number of seeds, number of pilot episodes and wall-clock inference budget per step.
Why not just…?
Why not just add a cost penalty to IDA's Q-function? This is the most direct threat. Penalty weighting is known to be fragile: low weights accept violations, and high weights give the "safe but useless" behavior documented by Li et al. and the oscillations reported in Safety Gymnasium. A penalty also ignores how much budget remains in the current episode. We run penalized IDA across a sweep of weights and plot its violation–intervention frontier against ours. If some fixed weight matches our frontier across pilots and tasks, the budget mechanism is unnecessary, and we will say so.
Why not just put a safety filter on the human's command? In open navigation with a single way around each hazard, a nearest-safe-action filter may perform as well as our copilot. The diffusion copilot should only matter where safe alternatives are multi-modal and the nearest safe action sits on a boundary between them. PACT reports that test-time guardrails underperform baked-in safety, with 31.0% fewer violations and 30.7% higher success for baked-in safety, but that comparison was autonomous manipulation, not shared control. We therefore include tasks deliberately built with symmetric hazards. We will report the filter as the preferred method on any task where it ties, since it is simpler.
Why not just discretize the commands, as DDS does? The 93% versus 26–56% closed-loop result is strong evidence that committing to a few options is what stabilizes feedback loops. Discretization, however, has no notion of cost and cannot represent fine approach trajectories near hazards. The DDS-plus-filter baseline and the mode-commitment ablation separate the two effects: stability from commitment, and safety from the budget. If DDS-plus-filter matches us on 2D tasks, our claim narrows to settings where continuous precision matters.
Why not just fix the decoder with domain adaptation offline? ADAN was evaluated offline and its authors note that users adapt in closed loop. FALCON found that unsupervised alignment failed to prevent degradation. Better decoding complements a copilot rather than replacing it. Our decoder-pilot experiments use a recalibrated decoder in all conditions, so this factor is held fixed.
Proposed Experiments and Decisive Checks
The benchmarks are Safety-Gymnasium PointGoal, CarGoal and PointButton at levels 1–2, using DSRL data for offline training, plus Lunar Lander with crashes and landing-pad overshoot counted as cost. We add two custom symmetric-hazard variants of PointGoal. Metrics are episode cost relative to $\epsilon$, task success, interventions per episode, action jerk, mode switches per intervention episode, and, for learning pilots, cumulative violations during training and final unassisted success.
Three checks decide the outcome. First, on symmetric-hazard tasks with decoder pilots, the copilot must beat both penalized IDA and the external filter on the violation–intervention frontier. Without mode commitment it should lose that advantage through chattering; if the ablation does not lose, the multi-modality argument fails. Second, for learning pilots, the copilot must keep cumulative violations under budget during training without lowering final unassisted success relative to the unassisted learner. If assistance stunts learning, that is itself a result. Third, we compute the rank correlation between open-loop replay scores and closed-loop scores across all copilots. We predict low correlation, with always-on methods favored in replay, which would show that copilot evaluations in this area must be run closed loop. Violations will be broken down by in-support versus out-of-support states, so that critic failure of the Safety Paradox kind is reported directly.
References
- Less is more: selection from a small set of options improves BCI velocity control
- Shared Autonomy with IDA: Interventional Diffusion Assistance
- To the Noise and Back: Diffusion for Shared Autonomy
- Safe Offline Reinforcement Learning with Feasibility-Guided Diffusion Model
- Don’t Trade Off Safety: Diffusion Regularization for Constrained Offline RL
- How Does the Lagrangian Guide Safe Reinforcement Learning through Diffusion Models?
- PACT: Self-Evolving Physical Safety Alignment for Diffusion Policies in Embodied Manipulation
- Breaking Safety Paradox with Feasible Dual Policy Iteration
- Neural Data Transformer 2: Multi-context Pretraining for Neural Spiking Activity
- Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark
- Datasets and Benchmarks for Offline Safe Reinforcement Learning
- Falcon: Fast Visuomotor Policies via Partial Denoising
- MetaDrive: Composing Diverse Driving Scenarios for Generalizable Reinforcement Learning
- Safety Gymnasium: A Unified Safe Reinforcement Learning Benchmark
- Adversarial Domain Adaptation for Stable Brain-Machine Interfaces