TacMan-Turbo: Proactive Tactile Control

• TacMan-Turbo is a tactile control framework that proactively uses contact deviations as informative kinematic signals rather than mere error corrections.
• It predicts the next handle pose with a constant-velocity model, merging progress and recovery phases for continuous, smooth task execution.
• Experimental results demonstrate dramatic improvements in time efficiency, action efficiency, and motion smoothness compared to reactive baseline methods.

TacMan-Turbo is a proactive tactile control framework for articulated object manipulation that treats tactile contact deviations not merely as errors to be corrected reactively, but as informative measurements of local kinematics. Introduced in "TacMan-Turbo: Proactive Tactile Control for Robust and Efficient Articulated Object Manipulation" (Zhao et al., 4 Aug 2025), it is designed to address a central difficulty in manipulating articulated objects in human environments: achieving both effectiveness, defined as reliable operation despite uncertain object structures, and efficiency, defined as swift execution with minimal redundant steps and smooth actions. The framework requires no pre-specified kinematics, is training-free, and is presented as compatible with geometry-based tactile sensors such as GelSight and also usable with force/torque sensors via calibration to geometry.

1. Problem setting and conceptual shift

Articulated object manipulation is framed as a setting in which robots must operate effectively under unknown or varying kinematic structure while also avoiding slow, redundant, or oscillatory behavior (Zhao et al., 4 Aug 2025). Traditional model-based methods rely on accurate prior kinematic models, such as revolute or prismatic joint types and parameters, and therefore lose effectiveness when those assumptions fail because of unknown structure or manufacturing variation. Prior tactile-informed methods avoid explicit kinematic priors, but the summary identifies a fundamental trade-off: robustness is obtained through reactive, step-by-step exploration-compensation cycles that compromise time efficiency, action efficiency, and trajectory smoothness.

TacMan-Turbo is defined by a different interpretation of tactile feedback. Rather than viewing contact deviations solely as error signals, it interprets them as rich sources of local kinematic information. This suggests a change in control semantics: tactile deformation is not only diagnostic of misalignment, but also predictive of how the handle is moving along its constrained path. On that basis, the controller estimates in-hand object pose from contact geometry, predicts the next handle pose with a constant-velocity model, and proactively computes an offset end-effector velocity that keeps the gripper aligned with the object’s constrained path. The reported consequence is continuous task progress without stop-and-go “explore-then-recover” cycles.

The contrast with Tac-Man, the recent tactile-informed baseline identified in the summary, is central. Tac-Man advances, accumulates deviation, and then pauses to compensate. TacMan-Turbo instead uses sequential contact deviations to extract local kinematic information and predict the next handle pose, computing a proactive velocity that avoids separate correction phases. The framework is therefore presented as resolving the long-standing effectiveness-efficiency trade-off without relying on prior kinematic knowledge.

2. State representation and proactive control formulation

The control formulation is expressed on $SE(3)$ with $4\times4$ homogeneous transforms $T$, rotation $R \in SO(3)$, and position $p \in \mathbb{R}^3$ (Zhao et al., 4 Aug 2025). At step $i$, the gripper and handle poses are $T_{g_i}$ and $T_{h_i}$, and initialization aligns them:

$T_{h_1} = T_{g_1}.$

The contact state at step $i$ is the set of active contact points sensed in the gripper frame,

$4\times4$0

with the more general dependence

$4\times4$1

The base end-effector twist in the gripper frame is

$4\times4$2

and an offset twist $4\times4$3 is computed so that the commanded twist becomes $4\times4$4.

The ideal proactive objective is to choose an offset twist that minimizes deviation of the next-step contact $4\times4$5 from the reference $4\times4$6, using a correspondence set $4\times4$7, while satisfying elementwise velocity limits. The summary states that the practical realization of this objective is achieved by setting the desired next gripper pose equal to the predicted next handle pose,

$4\times4$8

which ideally drives the next-step contact deviation toward zero. In this formulation, anticipation is built directly into the control target: the gripper is commanded not toward the current estimate of the handle pose, but toward its predicted next pose.

This proactive formulation differs from reactive compensation in a precise way. Reactive control waits for deviation to accumulate and then nulls it. TacMan-Turbo embeds anticipated deviation into the commanded motion itself. A plausible implication is that the controller does not separate “progress” and “recovery” into different phases; instead, both are merged into one continuously updated twist command.

3. Contact deviations as local kinematic information

The framework’s key estimation step is the recovery of the relative transform from gripper to handle from tactile contact geometry (Zhao et al., 4 Aug 2025). Let $4\times4$9 denote the unknown relative transform from gripper to handle at step $T$0, with $T$1. Under the assumption of no slip in the contact patch, $T$2 is estimated from correspondences between $T$3 and $T$4 by solving

$T$5

The summary identifies this as the standard point-set registration problem and states that it is solved efficiently by the Kabsch algorithm. It also gives an explicit geometric condition: at least three non-collinear contact points in $T$6 are required. The corresponding handle pose estimate is

$T$7

The interpretation of contact deviations is geometric. When the gripper moves relative to a constrained handle, the contact patch deforms and the set of $T$8D contact points shifts in the gripper frame. These displacements are described as a geometric “fingerprint” of the relative rigid transform between handle and gripper. By registering the current contact points to the initial reference contact set, the controller extracts a local estimate of handle motion without an explicit joint model.

The next handle pose is then predicted using a constant-velocity model over short intervals and smooth motions:

$T$9

$R \in SO(3)$0

The summary states that residual errors are corrected at the next step by re-estimation, preventing drift, and that the model can be extended to constant-acceleration if needed by incorporating more history. The local-motion estimate $R \in SO(3)$1 is thus treated as an approximation to the local kinematic motion, whether along a revolute arc, a prismatic line, or a more complex curve. This is the paper’s principal theoretical claim: contact deviations supply instantaneous motion constraints without requiring an explicit joint model.

4. Offset twist computation and execution constraints

Once the desired next gripper pose is fixed as the predicted next handle pose, the controller computes the transform needed in the current gripper frame (Zhao et al., 4 Aug 2025):

$R \in SO(3)$2

The ideal linear component of the offset twist is

$R \in SO(3)$3

For the angular component, with

$R \in SO(3)$4

and $R \in SO(3)$5 the unit rotation axis extracted from $R \in SO(3)$6, the controller uses

$R \in SO(3)$7

The ideal offset twist is therefore

$R \in SO(3)$8

Execution is constrained by robot joint limits. With robot joints $R \in SO(3)$9 and Jacobian $p \in \mathbb{R}^3$0, the induced joint rates are

$p \in \mathbb{R}^3$1

where $p \in \mathbb{R}^3$2 denotes the Moore–Penrose pseudoinverse. If $p \in \mathbb{R}^3$3 exceeds the joint limits $p \in \mathbb{R}^3$4, the controller scales with $p \in \mathbb{R}^3$5:

$p \in \mathbb{R}^3$6

The executed offset twist becomes

$p \in \mathbb{R}^3$7

and the execution time is stretched to

$p \in \mathbb{R}^3$8

The summary makes the operative assumptions explicit: non-slip contact over the sampling interval, at least three non-collinear contact points per step, constant-velocity prediction, and velocity or joint-limit compliance via scaling. These assumptions define the regime in which the method is expected to behave as reported. They also delimit failure modes: small or degenerate contacts, slip, rapid acceleration, or sensor-rate limitations can degrade estimation or prediction quality.

5. Experimental methodology and reported performance

The experimental program comprises both simulation and real-world evaluation (Zhao et al., 4 Aug 2025). In simulation, NVIDIA Isaac Sim runs at $p \in \mathbb{R}^3$9 Hz, and tactile feedback is modeled with the contact simulation method of Zhao et al. (2024) by displacements in contact regions. The robot is a $i$0-DoF Franka arm with full $i$1-DoF Cartesian capability. Robot placement is performed via B* to keep motions in workspace, and Cartesian control uses RMPFlow. Each trial starts from a stable grasp to define the reference contact $i$2.

The object set totals $i$3 articulated objects: $i$4 prismatic-joint objects from PartNet-Mobility with linear trajectories, $i$5 revolute-joint objects from PartNet-Mobility with circular trajectories, and $i$6 custom “complex” objects with $i$7D Bézier-curve trajectories of orders $i$8, $i$9, $T_{g_i}$0, and $T_{g_i}$1, with $T_{g_i}$2 in each order, mapped to a $T_{g_i}$3D playboard with a toy train serving as the handle. The curves are sampled and constrained to avoid self-intersections. The stated purpose of this set is to stress non-linear, variable-curvature paths.

The baseline is Tac-Man (Zhao et al., 2024), and both methods share identical initialization, grasping, and initial direction to isolate algorithmic differences. Success is defined differently by category: revolute rotation greater than $T_{g_i}$4, prismatic travel greater than $T_{g_i}$5 mm, and Bézier trajectories reaching the far end of the latent coordinate $T_{g_i}$6, with

$T_{g_i}$7

The evaluation metrics are success rate, time efficiency, action efficiency, and smoothness. Time efficiency is total completion time for successful trials. Action efficiency is the fraction of motion that contributes directly to progress along the task’s latent coordinate. Smoothness is defined as time-weighted joint-angle jerk, that is, the rate of change of acceleration scaled by task time, to capture motion quality without incentivizing trivially slow motions.

In the real world, the platform is a $T_{g_i}$8-DoF Kinova Gen3 with a Robotiq 2F-85 gripper. Two GelSight-type tactile sensors with an $T_{g_i}$9 marker grid on each finger replace the stock pads. Tactile update is $T_{h_i}$0 Hz, offset velocity is computed at $T_{h_i}$1 Hz, and low-level control runs at $T_{h_i}$2 Hz. The objects are a drawer, a microwave oven, and a bench vise with a helical handle path. The protocol uses identical initial stable grasp and initial direction for both methods, and the maximum gripper force is kept below the tactile sensor’s tolerance to satisfy non-slip. Total time and $T_{h_i}$3D gripper trajectories are recorded, and tactile force or geometry is visualized during motion.

6. Quantitative findings, implications, and limitations

Across the $T_{h_i}$4 simulated articulated objects, both TacMan-Turbo and Tac-Man achieved a $T_{h_i}$5 success rate for prismatic, revolute, and Bézier trajectories of orders $T_{h_i}$6 through $T_{h_i}$7 (Zhao et al., 4 Aug 2025). The distinction therefore lies not in nominal task completion but in the reported efficiency and smoothness of the manipulation.

Metric	TacMan-Turbo	Tac-Man
Success rate	100%	100%
Time efficiency	$T_{h_i}$8 s, $T_{h_i}$9 s	$T_{h_1} = T_{g_1}.$0 s, $T_{h_1} = T_{g_1}.$1 s
Action efficiency	$T_{h_1} = T_{g_1}.$2, $T_{h_1} = T_{g_1}.$3	$T_{h_1} = T_{g_1}.$4, $T_{h_1} = T_{g_1}.$5
Smoothness	$T_{h_1} = T_{g_1}.$6, $T_{h_1} = T_{g_1}.$7	$T_{h_1} = T_{g_1}.$8, $T_{h_1} = T_{g_1}.$9

The statistical tests reported in the summary are correspondingly strong. For time efficiency, the paper reports $i$0 and $i$1, with box plots on log-scale seconds showing consistently shorter completion times across every category and all $i$2. For action efficiency, the reported values are $i$3 and $i$4. For stability, measured as relative standard deviation, TacMan-Turbo has $i$5, $i$6, while Tac-Man has $i$7, $i$8, with $i$9 and $4\times4$00. For smoothness, measured by time-weighted joint jerk, the reported test is $4\times4$01 and $4\times4$02. The summary further states that the benefit is most dramatic on prismatic and revolute tasks, while improvements remain significant on Bézier trajectories despite occasional prediction challenges near inflection points.

The real-world results are qualitative and timing-based. On the drawer, microwave, and vise, both methods succeed, but TacMan-Turbo is reported to exhibit more direct, smoother trajectories and to complete tasks substantially faster. The teaser reports roughly $4\times4$03 higher time efficiency on the drawer, $4\times4$04 on the microwave, and $4\times4$05 on the vise, measured by manipulation speed. Across simulation and physical experiments, all reported efficiency improvements are stated to be highly significant with $4\times4$06. The summary also states that TacMan-Turbo maintains progress under disturbances without stopping for separate “recovery,” unlike reactive methods.

The advantages identified in the summary are explicit: the framework maintains $4\times4$07 success without kinematic priors while improving time efficiency by orders of magnitude, increasing action efficiency, and reducing jerk; it is training-free and model-free; it generalizes across linear, rotational, and complex paths; it eliminates stop-and-go recovery phases; it is sensor-agnostic in principle; and it is demonstrated on household objects under sensor noise and execution uncertainty. The limitations are equally explicit: accurate pose estimation requires non-slip and sufficient contact richness, especially at least three non-collinear points; the constant-velocity model can temporarily mispredict at inflection points or under rapid accelerations; the initial base velocity $4\times4$08 affects performance, with misalignment degrading action and time efficiency and excessive magnitude risking slip or sensor-sampling violations; and maximum usable speed is constrained by tactile sampling frequency and slippage threshold, with vision-based tactile sensors around $4\times4$09 Hz limiting peak speed more than high-rate capacitive or $4\times4$10-axis force/torque sensors above $4\times4$11 kHz.

The future extensions proposed in the summary follow directly from those limitations: adaptive base-velocity selection, higher-order predictors such as constant-acceleration or learning-augmented variants, broader tactile modalities and higher rates, multi-contact and more complex mechanisms, and vision–tactile fusion for coarse directional priors combined with tactile local kinematics. This suggests that TacMan-Turbo is best understood not only as a single controller, but as a control principle: tactile deviations are recast as local kinematic signal, and proactive alignment is built from that signal rather than from prior articulated-object models.