Test-Time Adaptation of Manipulation Policies Under Actuator Degradation

Som Sagar; Ransalu Senanayake

cs.ROSep 28, 2026

Test-Time Adaptation of Manipulation Policies Under Actuator Degradation

Authors: Som Sagar, Ransalu Senanayake

Organizations: School of Computing and Augmented Intelligence, Arizona State University, Tempe, Arizona

Abstract

Robot manipulation policies are usually trained under the assumption that a commanded action produces the same motion as it did during training even after hours of operation. Real hardware violates this assumption as the motors gradually heat up, current saturates near contact, voltage sags under load, thus the same policy action can produce a weaker, delayed, or noisier motion. These conditions are already measured by onboard telemetry, such as joint temperature, motor current, and supply voltage, yet this signal is typically used only for logging or safety checks rather than policy adaptation. We introduce Telemetry-Aware Action Rectification (TeAR), a policy-agnostic method that turns a frozen manipulation policy into a telemetry-conditioned policy by rectifying its outgoing action before it reaches the low-level controller. TeAR learns a lightweight Transformer that combines the proposed action with live actuator telemetry and amplifies, damps, or biases individual action components. We evaluate TeAR across 18 policy-task pairs spanning 8 policy families and 5 manipulation tasks. In an additional paired evaluation with degradation-model mismatch, TeAR achieves 31.8% success, compared with 25.6% for the base policy and 30.6% for an assumed-model inverse. On a physical arm, TeAR improves success under heating by 10-15% without on-robot fine-tuning.

Figures & tables

Fig. 1 : (a) Mean stressed success rate for 7 base policies ( $2$ M to $7$ B parameters): TeAR (hatched) adds $+12\%$ on average over the frozen base (solid). (b) Over one stressed episode, TeAR (red) corrects the base intent (blue) across axes; the inset shows the rectification $a_{b}\!\to\!a_{f}$ on a SO-101 arm.

Fig. 2 : Telemetry-Aware Action Rectification. A policy assumes its command $a_{b}$ and executes: on healthy hardware it succeeds (left), but under joint heating, or voltage sag the executed action $a_{\text{exec}}$ drifts from $a_{b}$ and the task fails (middle). TeAR (right) reads per-joint temperature, current, and voltage and rewrites each command into $a_{f}$ so execution recovers.

Fig. 3 : A frozen policy proposes an action vector. Each action component is paired with indexed actuator telemetry to form a telemetry–action token. Together with a robot-state token, these inputs are processed by a three-layer Transformer. Two output heads predict multiplicative and additive corrections, activated by a telemetry-dependent gate.

Fig. 4 : Degradation model. Per-joint capacity factor $\rho\in[\rho_{\min},1]$ for each telemetry channel: (a) temperature reduces torque capacity, (b) current induces saturation and cogging ripple, and (c) voltage limits attainable velocity. Shaded bands mark the physical operating regimes.

Task	Base policy	cool	hot	stall	brown	$\mathcal{T}_{\text{mod}}$	$\mathcal{TC}_{\text{mod}}$	$\mathcal{TV}_{\text{mod}}$	$\mathcal{TCV}_{\text{mod}}$	$\Delta(\uparrow)\pm\sigma$
Lift	BC ( $\sim$ 2M)	85	40 $\rightarrow$ 65	50 $\rightarrow$ 55	45 $\rightarrow$ 40	75 $\rightarrow$ 90	80 $\rightarrow$ 85	80 $\rightarrow$ 100	85 $\rightarrow$ 90	$+$ 7.2 $\pm$ 1.4
	BC-Transformer (18M)	99	59 $\rightarrow$ 78	36 $\rightarrow$ 50	67 $\rightarrow$ 61	93 $\rightarrow$ 90	84 $\rightarrow$ 94	87 $\rightarrow$ 93	69 $\rightarrow$ 75	$+$ 6.8 $\pm$ 0.2
	ACT [ 1 ] ( $\sim$ 52M)	100	57 $\rightarrow$ 90	27 $\rightarrow$ 71	44 $\rightarrow$ 63	99 $\rightarrow$ 100	93 $\rightarrow$ 100	87 $\rightarrow$ 100	56 $\rightarrow$ 96	$+$ 22.2 $\pm$ 1.4
	GR00T-N1.5 [ 24 ] (3B)	100	40 $\rightarrow$ 78	13 $\rightarrow$ 54	17 $\rightarrow$ 81	93 $\rightarrow$ 99	61 $\rightarrow$ 100	49 $\rightarrow$ 97	16 $\rightarrow$ 92	$+$ 44.8 $\pm$ 2.3
	OpenVLA [ 11 ] (7B)	100	50 $\rightarrow$ 65	10 $\rightarrow$ 55	40 $\rightarrow$ 40	100 $\rightarrow$ 100	96 $\rightarrow$ 98	100 $\rightarrow$ 100	95 $\rightarrow$ 98	$+$ 9.3 $\pm$ 2.5
Can	BC ( $\sim$ 2M)	39	9 $\rightarrow$ 11	2 $\rightarrow$ 0	4 $\rightarrow$ 0	13 $\rightarrow$ 35	15 $\rightarrow$ 24	16 $\rightarrow$ 43	3 $\rightarrow$ 0	$+$ 7.3 $\pm$ 2.9

TABLE I: Task success (%) for 18 policy–task pairs across eight policy families and five tasks. Stressed-condition cells report base

\rightarrow

TeAR success; cool reports nominal success. Hot, stall, and brown apply temperature, current, and voltage stress individually; the moderate profiles activate the channels indicated by their subscripts.

\Delta

reports the mean stressed-condition gain over the base, with the reported standard deviation across evaluation seeds.

Fig. 5 : Environment setup for real (left-2) and sim (right-2).

Fig. 6 : Stressed-condition

\Delta

over the frozen base for five recipes with the TeAR architecture fixed.

Method	hot	$\mathcal{T}_{\text{mod}}$	$\mathcal{TC}_{\text{mod}}$	$\mathcal{TCV}_{\text{mod}}$	$\Delta(\uparrow)$
Frozen base	$6.7\,{\scriptscriptstyle\pm 2.9}$	$43.3\,{\scriptscriptstyle\pm 12.6}$	$26.7\,{\scriptscriptstyle\pm 5.8}$	$0.0\,{\scriptscriptstyle\pm 0.0}$	—
No-Tele. SFT	$3.3\,{\scriptscriptstyle\pm 5.8}$	$41.7\,{\scriptscriptstyle\pm 10.4}$	$0.0\,{\scriptscriptstyle\pm 0.0}$	$0.0\,{\scriptscriptstyle\pm 0.0}$	$-6.9\,{\scriptscriptstyle\pm 3.4}$
Domain R. [ 19 ]	$5.0\,{\scriptscriptstyle\pm 0.0}$	$46.7\,{\scriptscriptstyle\pm 11.5}$	$25.0\,{\scriptscriptstyle\pm 13.2}$	$0.0\,{\scriptscriptstyle\pm 0.0}$	$0.0\,{\scriptscriptstyle\pm 3.1}$
ACDR [ 22 ]	$8.3\,{\scriptscriptstyle\pm 2.9}$	$40.0\,{\scriptscriptstyle\pm 10.0}$	$31.7\,{\scriptscriptstyle\pm 7.6}$	$0.0\,{\scriptscriptstyle\pm 0.0}$	$+1.0\,{\scriptscriptstyle\pm 3.2}$
RMA-style [ 15 ]	$5.0\,{\scriptscriptstyle\pm 5.0}$	$35.0\,{\scriptscriptstyle\pm 13.2}$	$23.3\,{\scriptscriptstyle\pm 5.8}$	$0.0\,{\scriptscriptstyle\pm 0.0}$	$-4.3\,{\scriptscriptstyle\pm 2.6}$
BC fine-tune	$5.0\,{\scriptscriptstyle\pm 0.0}$	$46.7\,{\scriptscriptstyle\pm 11.5}$	$25.0\,{\scriptscriptstyle\pm 13.2}$	$0.0\,{\scriptscriptstyle\pm 0.0}$	$0.0\,{\scriptscriptstyle\pm 3.1}$

TABLE II : Baselines. Success (%), mean

\pm

sample SD across 3 evaluation seeds.

\Delta

averages paired gains over the frozen BC-Transformer across seven stressed conditions.

Fig. 7 : Stressed success gain sweeping the additive-residual cap $\alpha$ and the multiplicative gain $\gamma_{\max}$ . The selected $(\alpha,\gamma_{\max})=(0.30,0.75)$ is outlined.

Configuration	$\mathcal{T}_{\text{mod}}(\uparrow)$	$\mathcal{TV}_{\text{mod}}(\uparrow)$	Stressed( $\uparrow$ )	$\Delta_{T,TV}$
TeAR (reference)	$63.3\,{\scriptscriptstyle\pm 2.9}$	$68.3\,{\scriptscriptstyle\pm 12.6}$	$34.8\,{\scriptscriptstyle\pm 0.4}$	$0.0\,{\scriptscriptstyle\pm 0.0}$
Gate $\to$ smoothstep	$53.3\,{\scriptscriptstyle\pm 10.4}$	$48.3\,{\scriptscriptstyle\pm 16.1}$	$34.3\,{\scriptscriptstyle\pm 3.1}$	$-15.0\,{\scriptscriptstyle\pm 13.2}$
Output: $\delta$ only	$46.7\,{\scriptscriptstyle\pm 5.8}$	$58.3\,{\scriptscriptstyle\pm 15.3}$	$31.7\,{\scriptscriptstyle\pm 4.2}$	$-13.3\,{\scriptscriptstyle\pm 5.2}$
Longer SFT ( $2\times$ )	$61.7\,{\scriptscriptstyle\pm 12.6}$	$58.3\,{\scriptscriptstyle\pm 2.9}$	$33.8\,{\scriptscriptstyle\pm 2.9}$	$-5.8\,{\scriptscriptstyle\pm 11.5}$
Gate $\to$ linear	$55.0\,{\scriptscriptstyle\pm 5.0}$	$50.0\,{\scriptscriptstyle\pm 10.0}$	$33.1\,{\scriptscriptstyle\pm 2.2}$	$-13.3\,{\scriptscriptstyle\pm 7.6}$
Output: $\gamma$ only	$53.3\,{\scriptscriptstyle\pm 11.5}$	$51.7\,{\scriptscriptstyle\pm 7.6}$	$36.0\,{\scriptscriptstyle\pm 3.9}$	$-13.3\,{\scriptscriptstyle\pm 12.6}$

TABLE III: Success (%) as mean

\pm

sample SD across three evaluation seeds with fixed checkpoints. Stressed averages all seven stress conditions;

\Delta_{T,TV}

is the paired gain over full TeAR averaged over

\mathcal{T}_{\text{mod}}

and

\mathcal{TV}_{\text{mod}}

conditions.

Adapter	hot	$\mathcal{T}_{\text{mod}}$	$\mathcal{TC}_{\text{mod}}$	$\mathcal{TCV}_{\text{mod}}$	$\Delta(\uparrow)$
Frozen BC-T	$6.7\,{\scriptscriptstyle\pm 2.9}$	$43.3\,{\scriptscriptstyle\pm 12.6}$	$26.7\,{\scriptscriptstyle\pm 5.8}$	$0.0\,{\scriptscriptstyle\pm 0.0}$	—
Bottleneck res	$10.0\,{\scriptscriptstyle\pm 0.0}$	$51.7\,{\scriptscriptstyle\pm 11.5}$	$1.7\,{\scriptscriptstyle\pm 2.9}$	$10.0\,{\scriptscriptstyle\pm 5.0}$	$+1.9\,{\scriptscriptstyle\pm 3.5}$
Cross-attn	$15.0\,{\scriptscriptstyle\pm 10.0}$	$55.0\,{\scriptscriptstyle\pm 5.0}$	$10.0\,{\scriptscriptstyle\pm 5.0}$	$13.3\,{\scriptscriptstyle\pm 10.4}$	$+6.0\,{\scriptscriptstyle\pm 6.8}$
FiLM	$28.3\,{\scriptscriptstyle\pm 2.9}$	$20.0\,{\scriptscriptstyle\pm 5.0}$	$13.3\,{\scriptscriptstyle\pm 5.8}$	$8.3\,{\scriptscriptstyle\pm 5.8}$	$+1.0\,{\scriptscriptstyle\pm 2.3}$
Per-joint MLP	$15.0\,{\scriptscriptstyle\pm 5.0}$	$45.0\,{\scriptscriptstyle\pm 13.2}$	$3.3\,{\scriptscriptstyle\pm 2.9}$	$6.7\,{\scriptscriptstyle\pm 5.8}$	$0.0\,{\scriptscriptstyle\pm 5.0}$
Global MLP	$23.3\,{\scriptscriptstyle\pm 2.9}$	$36.7\,{\scriptscriptstyle\pm 16.1}$	$8.3\,{\scriptscriptstyle\pm 7.6}$	$10.0\,{\scriptscriptstyle\pm 0.0}$	$+1.2\,{\scriptscriptstyle\pm 5.8}$

TABLE IV : Adapter architectures. Success (%), mean

\pm

sample SD across three evaluation seeds with fixed checkpoints, using the same SFT recipe and gate.

\Delta

averages gains over the frozen base across seven stressed conditions.

Fig. 8 : Action correction under combined temperature and current stress. Top: paired base-policy and TeAR snapshots from the same initial state; yellow circles mark the target can. TeAR completes the task at 11.95 s, while the base fails within 20 s. Bottom: proposed and corrected

y

-translation commands along the TeAR trajectory. Shading highlights amplification and damping, with numeric examples of each. This selected rollout illustrates correction behavior.

Fig. 9 : Compensation under model mismatch. Bars show corrected minus base success rates; positive values indicate improvement. Each task uses 480 episodes per method; Overall averages the four tasks.

Condition	Base	TeAR	$\Delta(\uparrow)$
Cool ( $<45^{\circ}$ C)	100	100	$\pm 0$
Hot (shoulder $>50^{\circ}$ C)	75	85	$+10$
Hot (shoulder & wrist $>50^{\circ}$ C)	70	85	$+15$

TABLE V: SO-101 sim-to-real. Lift success rates (%) over 20 trials per method and condition;

\Delta

is the % gain.

Fig. 10 : Per-channel capacity and perturbation magnitudes. The implemented capacities are linear ramps. Dashed lines show Gaussian standard deviations; the dotted current trace shows ripple amplitude. All noise and ripple terms vanish at full capacity.

Fig. 11 : Telemetry gate. Left: the combined temperature–current gate with nominal voltage; the marker lies in the exactly closed nominal region. Right: each channel uses the same smoothstep in its own normalized severity coordinate. Temperature gate corners are $42$ – $55^{\circ}$ C; current and voltage corners are $0.60$ – $1.00$ and $0.90$ – $0.50$ . Gate activation is distinct from the capacity ramps in Fig. 10 .

Trunk	Can	Threading	Stack	Mean
Reference	$12.9$	$35.0$	$62.9$	$36.9$
XL	$13.6$	$30.7$	$66.4$	$36.9$

TABLE VI : Adapter size comparison. Success (%) under the same seven-condition aggregation.

Tier	Temperature ( ∘ C)	Current	Voltage
Clean	$[20,42]$	$[0.10,0.50]$	$[0.92,1.00]$
Mild	$[43,55]$	$[0.50,0.75]$	$[0.80,0.92]$
Moderate	$[50,65]$	$[0.65,0.90]$	$[0.65,0.85]$
Severe	$[60,75]$	$[0.80,1.05]$	$[0.45,0.70]$

TABLE VII : Telemetry ranges for SFT augmentation. Current and voltage are normalized ratios. Each tier contributes equally many examples.

Fig. 12 : Physical SO-101 platform. (a) The arm used for cube lifting. (b) Thermal-camera spot readings. (c) The thermal image illustrates spatially nonuniform heating. The photographed readings illustrate the setup, rather than the heated evaluation thresholds.

SFT curriculum	Success (%)	$\Delta$ vs. four tiers
Clean only	$23.8$	$-13.1$
Clean + mild	$39.5$	$+2.6$
Clean + mild + moderate	$40.7$	$+3.8$
Four tiers (reference)	$36.9$	$0.0$

TABLE VIII : Severity coverage during training. Stressed success averaged over three cells and seven conditions. Differences are percentage points relative to the four-tier recipe.

Trunk	Success (%)	$\Delta$
Bottleneck	$26.8\pm 21.7$	$-8.2$
Cross-attention	$32.3\pm 24.4$	$-2.7$
FiLM	$28.2\pm 22.0$	$-6.8$
Hybrid	$26.0\pm 20.9$	$-9.0$
Per-index MLP	$29.3\pm 19.3$	$-5.7$
Global MLP	$28.8\pm 21.1$	$-6.2$

TABLE IX : Alternative adapter architectures. Mean success and sample standard deviation across six cells; changes are percentage points from the reference.

Fig. 13 : Selection of correction bounds. Mean stressed success-rate gain over the frozen base (percentage points), reproduced from the main paper. The outlined cell marks the selected configuration, $(\alpha,\gamma_{\max})=(0.30,0.75)$ .

Visible channels	Success (%)	$\Delta$ vs. full
T only	$28.3$	$-4.0\pm 4.3$
C only	$16.5$	$-15.8\pm 6.7$
V only	$19.1$	$-13.2\pm 7.8$
T + C	$31.1$	$-1.1\pm 5.3$
T + V	$28.8$	$-3.5\pm 4.4$
C + V	$18.6$	$-13.7\pm 9.6$

TABLE X : Telemetry available at inference. Mean stressed success over ten cells.

\Delta

is the paired change from full telemetry, reported as mean

\pm

sample standard deviation across cells, in percentage points.

Fig. 14 : Mean absolute correction by action coordinate. Values are recomputed from all ten saved cell summaries. The nominal row is exactly zero; combined stress produces larger corrections, especially in translation. The gripper may receive a learned correction even though the synthetic degradation leaves its command unchanged.

Fig. 15 : Evaluated capacity curves. Colored lines show the eight test curve sets; dashed black lines show the nominal training model. Each set changes all three channels jointly, varying curve shape, onset, and minimum capacity.

Explore similar work

Proxy Policy Steering
Sep 8, 2026Chuanruo Ning, Tianrui Wang, Wei-Chiu Ma +1Robot Policies Robot Systems
Cornell University
ORPA: Online Residual Policy Adaptation for Robot Manipulation Control with Human Feedback
Aug 18, 2026Muhammad A. Muttaqien, Tomohiro Motoda, Ryo Hanai +1Residual Policy Robotic Manipulation Policies
Embodied AI Research Team · National Institute of AIST · Tokyo, Japan
TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation
Jun 4, 2026Dongwon Son, Florian Shkurti, Jason Lee +3Torque Robotic Manipulation Policies
1KAIST · *Work done during an internship at the Allen Institute for AI. · 2Allen Institute for AI +2

Sep 8, 2026, cs.ROJ/K move · Enter open · S save

Proxy Policy Steering

Chuanruo Ning, Tianrui Wang, Wei-Chiu Ma, Kuan Fang

Cornell University

Robot Policies Robot Systems Steering Inference-Time

Open paper AskPDF arXiv

Generalist robot policies carry broad manipulation priors from large-scale data, but specializing them to a new task remains the deployment bottleneck. This requires eliciting task-specific behavior from limited demonstrations without degrading their broad capabilities. We introduce Proxy Policy Steering (PPS), an inference-time adaptation method that resolves this challenge by training two lightweight proxy policies whose calibrated velocity-space difference steers the frozen base sampler. A reference proxy models the frozen base's behavior on target-task observations, and a task proxy, initialized from the reference, captures how this behavior changes under task supervision. Their difference forms a calibrated velocity-space residual that steers the frozen base sampler at every denoising step. We identify the conditions under which this residual isolates the change induced by task supervision, and validate them empirically. Because the base is never directly modified, its broad capabilities remain available at inference, including behaviors such as recovery from failure that the demonstrations themselves do not exercise. Adaptation requires only forward velocity predictions from the base, making PPS lightweight to train and applicable even without access to the base's parameters. On 8 real-world and 4 simulation manipulation tasks, PPS lifts the state-of-the-art pi 0.5 base policy by 53% absolute success rate on average, with zero-to-one gains on tasks the base never solves, while preserving the base's broad capabilities. PPS outperforms LoRA fine-tuning, from-scratch specialists, residual policies, and prior inference-time steering methods.

Abstract

Explore similar work

Proxy Policy Steering

ORPA: Online Residual Policy Adaptation for Robot Manipulation Control with Human Feedback

TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation