A robot policy that reports "90% sure this grasp will work" is only useful if that number means the same thing on every robot it runs on. Robotics software already solves a version of this portability problem for commands. A hardware abstraction layer (HAL) is the translation layer that lets the same high-level program drive many different arms, bases, or drones by converting abstract intents such as "move the gripper 2 cm left" into motor-specific signals, and by converting raw sensor readings into standard state. Confidence estimates get no such treatment. The best recent calibrated failure predictor for vision-language-action models (VLAs, large networks that map camera images and a language instruction to motor commands) collapses when moved from simulation to a real WidowX arm and has to be retrained from scratch (Temporal Difference Calibration in Sequential Tasks). Every new body therefore needs a new, expensive campaign of labeled rollouts before the robot's confidence can be trusted.
We propose making the HAL carry the information calibration needs. Our hypothesis is that confidence transfers poorly because existing predictors fold two separate things into one score: whether the policy has the right intent, and whether this particular body can execute that intent. The HAL is the one component that already knows the second. We extend it to report a small set of normalized execution residuals upstream: how far the arm lags the commanded motion relative to its own nominal tracking ability, how often actuators saturate, and how much a safety filter had to alter the command. A failure predictor that reads policy features alongside these canonical residuals should then transfer to a new embodiment with far fewer target-robot rollouts than one built on policy features alone. The quantity we care about is the number of target-body rollouts needed to reach a fixed calibration quality.
Core Claim
Body-specific dynamics enter a VLA's success probability mainly through execution feasibility, and the HAL can express feasibility in units that do not depend on the platform. Suppose $e_t$ is the end-effector tracking error at step $t$ and $\bar e$ is the platform's nominal tracking-error envelope, measured once with a short scripted excitation routine that needs no task labels. Let $s_t$ be the fraction of control ticks in which any actuator hit its limit, and $c_t = \|v^{ref}_t - v^{safe}_t\| / \|v^{ref}_t\|$ the relative correction applied by a kinematic safety filter. The predictor is
$$\hat p_t = \sigma\big(f(h_t) + g(e_{1:t}/\bar e,\ s_{1:t},\ c_{1:t})\big),$$
where $h_t$ is the VLA's hidden state, $f$ is a small learned network over policy features, $g$ is a small learned network over the residual history of the trajectory executed so far, and $\sigma$ is the logistic function. The prediction concerns eventual task success. It is trained with the time-dependent Brier objective and temporal-difference targets of TDQC, so we reuse that paper's estimator unchanged and alter only its inputs. The claim being tested is that $g$, trained on source embodiments, can be reused on a new body after only its $\bar e$ normalizer is recomputed. Only $f$ and a two-parameter output recalibration would then need target data.
Research Program
The work has three parts. The first is engineering. We extend an existing robot-learning middleware with an upstream "feasibility channel" that streams the three residuals at control rate next to the normal state. LeRobot suits this because it already separates remote policy inference from local motor loops across eight hardware setups. Its local loop sees tracking and saturation directly, but it currently runs action chunks open-loop with no uncertainty signal at all. For the safety-correction residual we use the two-layer design of Model-Free Safety-Critical Control for Robotic Systems: a kinematic barrier-function filter that is independent of mass and inertia, which makes its correction magnitude comparable across bodies by construction. We use that paper as a component that produces a signal. We do not adapt its aggressiveness parameter, because coupling confidence to safety parameters is a separate question.
The second part is the transfer study. We train the predictor on source embodiments, move it to a held-out embodiment, and measure calibration as a function of how many labeled target rollouts are allowed (0, 10, 25, 50, 100, 200).
The third part is analysis. We decompose miscalibration on the target body into a component explained by the residuals and a remainder. If the remainder dominates, the hypothesis is wrong: transfer failure would come from perception shift, such as new camera viewpoints, rather than from dynamics. That would be a useful negative result for anyone hoping abstraction alone can fix calibration.
The proposal builds on three lines of VLA calibration work and departs from each in the same way. Confidence Calibration in VLA Models introduces prompt ensembles and per-axis ("action-wise") Platt scaling on LIBERO. It motivates this work, because it shows that different degrees of freedom calibrate differently. It also constrains the design, because its per-axis parameters are tied to one action parameterization and must be refit for a new body. SCALE and UAOR use the policy's internal entropy to change how it attends or decodes at test time. They use uncertainty as a control signal and do not try to make it calibrated or portable. TDQC supplies our estimator. All of these treat the robot as an end-to-end black box. None asks the hardware layer what it observed.
Baselines
The naive reference point is the VLA's raw action-token confidence, taken as the mean maximum-probability of the action tokens over the executed prefix and applied unchanged on the target robot. It shows how bad things are when nothing is done.
The strongest existing method is TDQC itself: a temporal-difference success predictor over VLA features with conformal thresholds. We run it on the same source data and give it the same target-rollout budgets for fine-tuning, with the same hyperparameter search budget, which we fix at 20 configurations per method on a source-only validation split. Its reported results are task-success numbers, not cross-body calibration numbers: guided action search lifted OpenVLA on LIBERO-10 from 40% to 55% success, and thresholding saved nearly half the extra compute. It reports transfer failure from LIBERO to WidowX only qualitatively. We found no reported result on Lacuna for cross-embodiment ECE of any VLA failure predictor, so our TDQC transfer curve will be the first such number. Beating TDQC means reaching the same ECE and failure-detection AUROC with at least 4× fewer target rollouts, or reaching a lower ECE at every budget of 50 rollouts or fewer.
The second strong competitor is action-wise Platt scaling refit on the target body. Its source paper reports that prompt ensembling cuts ECE by over 20% on average on LIBERO. We apply it on top of TDQC at matched budgets.
A cross-domain reference is a hidden-state probe in the style of PROBECAL, which reduced ECE from 0.690 to 0.035 on TabMWP tool use using 500 training examples. That number comes from text tasks and is not directly comparable to robot rollouts.
Two ablations isolate the key ingredient. The first gives the predictor the same residuals without the $\bar e$ normalization. If transfer still works, normalization is not the reason. The second uses residuals only, with no policy features, which tells us how much of success prediction is purely about the body.
Why not just…?
Why not just recalibrate on the new robot with a few dozen rollouts? This is the most serious threat. If two-parameter Platt scaling on 25 target rollouts recovers TDQC's source-level ECE, the HAL channel is unnecessary. The transfer curve settles this directly, since every method receives the same budgets. TDQC's own report that sim-to-WidowX transfer required complete retraining suggests the shift is more than a temperature change. We will report honestly wherever the curves cross. If they cross below 25 rollouts, we will drop the claim.
Why not just train the predictor on pooled data from many embodiments? Pooling could teach $f$ an embodiment-invariant notion of success without any HAL changes. We include a pooled-source TDQC baseline trained on exactly the same source bodies as our method. The difference between it and our method measures what explicit, normalized feasibility adds beyond data diversity. Pooling cannot help with a body whose tracking envelope lies outside the source range. Our held-out bodies are chosen to include one such case, a slower low-cost SO-100 arm.
Why not just probe the VLA's hidden states, which are already body-agnostic? The VLA's features describe what it sees and intends. They do not describe whether a weaker motor lagged, saturated, or was overridden by a safety filter. PROBECAL worked because digital tools execute deterministically. Physical execution is not deterministic, and the policy-features-only ablation measures exactly this gap.
Why not threshold entropy as SCALE does? SCALE's self-uncertainty improves success, from 52.7% to 63.3% on LIBERO-Long. However, it is a heuristic gate and was not evaluated as a calibrated probability. We include it as a scored baseline using its dual-reference uncertainty.
Proposed Experiments
In simulation, the source is LIBERO (LIBERO) with OpenVLA. For cross-body transfer we use the CoppeliaSim Franka-to-UR5/Sawyer setup and the Genesis gripper swaps (parallel jaw, Robotiq 2F-85, suction) from RECENT. These isolate kinematic and end-effector changes while holding perception roughly fixed. On real hardware we transfer from a Franka predictor to a WidowX following TDQC's 8-task pick-and-place protocol, and to an SO-100 running through LeRobot.
The metrics are ECE, Brier score, and failure-detection AUROC computed at 25%, 50%, and 75% of trajectory progress, because calibration is known to vary with progress. Each is reported against the target-rollout budget. We also measure a downstream outcome: success-weighted task throughput when low confidence triggers an abort and re-plan.
Three results are decisive. The first is the zero-shot point: with only $\bar e$ recomputed, our predictor should already beat TDQC at 0 target rollouts. The second is the 25-rollout comparison against target-refit Platt scaling. The third is the unnormalized-residual ablation, which should transfer markedly worse if normalization is doing the work we claim. If the first and third fail, we will conclude that the HAL cannot carry calibration and that the dominant shift is perceptual.
References
- Temporal Difference Calibration in Sequential Tasks: Application to Vision-Language-Action Models
- LeRobot: An Open-Source Library for End-to-End Robot Learning
- Model-Free Safety-Critical Control for Robotic Systems
- Confidence Calibration in Vision-Language-Action Models
- SCALE: Self-uncertainty Conditioned Adaptive Looking and Execution for Vision-Language-Action Models
- UAOR: Uncertainty-aware Observation Reinjection for Vision-Language-Action Models
- Uncertainty Calibration for Tool-Using Language Agents
- LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning
- Efficient Skill Grounding via Code Refactoring with Small Language Models