---
title: 'TATIC: Task-Aware Temporal Intent Inference'
url: https://www.emergentmind.com/topics/tatic
type: topic
---

# TATIC: Task-Aware Temporal Intent Inference

to=arxiv_search.search 下载彩神争霸াত্র json
{"query":"TATIC arXiv human robot collaboration physical corrections intent inference 2026", "max_results": 5}
to=arxiv_search.search  全民彩票天天送json
{"query":"\"Task-Aware Temporal Learning for Human Intent Inference from Physical Corrections in Human-Robot Collaboration\"", "max_results": 5}
to=arxiv_search.search  สล็อตโjson
{"query":"TATIC TacSIm Tactical Style Imitation football arXiv", "max_results": 10}
TATIC most directly denotes **Task-Aware Temporal Learning for Human Intent Inference from Physical Corrections in Human-Robot Collaboration**, a framework for decoding brief physical corrections in human-robot collaboration into both discrete task-level intent and continuous motion-level parameters. It combines torque-based contact force estimation, task-aligned feature canonicalization, a causal task-aware Temporal Convolutional Network (TCN), and an intent-driven adaptation layer that maps inferred semantics into path deformation, operation-space inflation, speed scaling, goal switching, or software-level hold. The formulation is motivated by contact-rich collaborative settings in which operators convey motion-level adjustments through short physical interactions, while the robot must recover task-level semantics under dynamic constraints and low-latency requirements [2603.11077].

## 1. Scope, intent vocabulary, and problem setting

TATIC addresses a specific gap in human-robot collaboration: existing foundation-model-based approaches primarily rely on vision and language inputs and lack mechanisms to interpret physical feedback, while traditional physical human-robot interaction methods use corrections for trajectory guidance but struggle to infer task-level semantics from brief contacts [2603.11077]. The core problem is therefore not only motion adaptation, but the joint recovery of **task-level intent** and **operator-conditioned continuous parameters** from contact segments whose durations were brief, specifically **0.53–1.98 s**.

The discrete intent vocabulary is fixed to
\[
o[t] \in \{\mathrm{Guide}, \mathrm{Yield}, \mathrm{Slow}, \mathrm{Stop}, \mathrm{Switch}\}.
\]
TATIC jointly estimates continuous quantities conditioned on the inferred operator: for **Guide**, a direction \(d \in S^2\) and magnitude \(m \in [0,1]\); for **Yield**, a safety-margin radius \(r_s\); for **Slow**, a speed scaling \(\alpha_v \in [\alpha_{\min}, \alpha_{\max}]\); and for **Switch**, a target index \(k \in \{0,1\}\) [2603.11077]. This design separates semantic intent from low-level execution variables while keeping the latter explicitly actionable.

Several challenges are explicit in the formulation. First, contact segments are short, so inference must be causal and low-latency. Second, task constraints such as task-centric safety zones and planner ambiguity vary online. Third, workspace reconfigurations change global pose, so a representation tied to world coordinates degrades generalization. TATIC addresses this by constructing a task-aligned canonical frame and learning over canonicalized temporal feature streams rather than raw global geometry. A plausible implication is that the framework is best understood as a **task-conditioned temporal inference system** rather than a generic force-to-action mapper.

## 2. Torque-based contact estimation and contact-segment formation

The sensing pipeline begins with measured joint torques \(\tau_{\mathrm{meas}}[k]\), together with proprioceptive variables \(q[k]\), \(\dot q[k]\), and \(\ddot q[k]\). The dynamics model is
\[
\tau = M(q)\ddot q + C(q,\dot q)\dot q + g(q) + J(q)^\top f_{\mathrm{ext}},
\]
and the formulation used in TATIC collects Coriolis/centripetal, gravity, and friction into \(h(q,\dot q)\), giving
\[
M(q)\ddot q + h(q,\dot q) + \tau_{\mathrm{ext}} = \tau_{\mathrm{meas}}.
\]
External torque from a contact at \(p_c\) is modeled as
\[
\tau_{\mathrm{ext}} = J_c(q,p_c)^\top f_c.
\]
The residual torque estimate is then
\[
\hat{\tau}_{\mathrm{ext}}[k] = \tau_{\mathrm{meas}}[k] - \left(M(q[k])\ddot q[k] + h(q[k],\dot q[k])\right),
\]
and the contact detection statistic is
\[
\eta[k] = \|W_\tau \hat{\tau}_{\mathrm{ext}}[k]\|_2,
\]
combined with EWMA filtering, a threshold \(\theta_\tau\), and persistence logic \(N_{\mathrm{on}}/N_{\mathrm{off}}\) for onset and offset detection [2603.11077].

Once contact is detected, TATIC localizes the contacted link coarsely by thresholding active joint residuals and choosing the highest active joint satisfying a persistence condition. Contact position is parameterized by a normalized scalar \(s \in [0,1]\) along the contacted link centerline. Over a short window of samples \(\{k_r\}_{r=1}^N\), the residual model is
\[
r_r(s,f) = \hat{\tau}_{\mathrm{ext}}[k_r] - J_c(q[k_r], p_c(s,k_r))^\top f,
\]
and the contact state is estimated by solving
\[
[s[t], f[t]] =
\arg\min_{s,f}\;
\frac{1}{2}\sum_{r=1}^N \|r_r(s,f)\|_2^2
\quad
\text{s.t. } 0 \le s \le 1,\; \|f\|_2 \le F_{\max}.
\]
For fixed \(s\), \(f\) is obtained by Tikhonov-regularized least-squares and projected onto the ball \(\|f\|\le F_{\max}\); \(s\) is searched on a uniform grid and refined via Brent search [2603.11077].

This stage is central because all downstream semantics are conditioned on the estimated contact geometry and force. The framework therefore does not treat physical correction as an abstract symbol stream; it explicitly reconstructs a physically grounded interaction state from residual dynamics.

## 3. Task-aligned feature canonicalization

A defining element of TATIC is **task-aligned feature canonicalization**, introduced to improve robustness under layout reconfiguration. Rather than learning directly in the world frame, the method constructs a local canonical frame \(E[t] \in SO(3)\) aligned with the planner-provided reference velocity \(v_{\mathrm{ref}}[t]\) and the world vertical axis \(\hat z\) [2603.11077].

The forward axis is
\[
e_1[t] =
\frac{v_{\mathrm{ref}}[t]}{\|v_{\mathrm{ref}}[t]\|}
\quad \text{if } \|v_{\mathrm{ref}}[t]\|>\epsilon_{\mathrm{ref}},
\]
otherwise \(e_1[t-1]\) is reused. An auxiliary axis is selected as
\[
a[t] = \hat z
\quad \text{if } \|\hat z \times e_1[t]\|>\epsilon_\times,
\]
otherwise \(a[t]=R_{ee}[t]e_y\). The remaining basis vectors are
\[
e_2[t]=\frac{a[t]\times e_1[t]}{\|a[t]\times e_1[t]\|}, \qquad
e_3[t]=e_1[t]\times e_2[t],
\]
yielding the raw frame \(\tilde E[t]=[e_1\ e_2\ e_3]\). Temporal smoothing on \(SO(3)\) is performed via
\[
E[t] = E[t-1]\exp\!\left(\lambda_E \log(E[t-1]^\top \tilde E[t])\right).
\]

Positions and vectors are then projected into the task frame. For positions,
\[
x_c = R_{\mathrm{task}}^\top (x - t_{\mathrm{task}}),
\]
and for general vectors,
\[
\mathrm{proj}_t(v)=E[t]^\top v.
\]
The canonicalized feature set includes kinematics, workspace safety, goal alignment, localization, and planner context [2603.11077]. Representative terms are the reference-speed magnitude \(v_{\mathrm{mag}}=\|v_{\mathrm{ref}}\|\); local unit force direction
\[
\hat f_{\mathrm{loc}}[t]=\frac{\mathrm{proj}_t(f[t])}{\|\mathrm{proj}_t(f[t])\|};
\]
workspace clearance
\[
d_{hw}[t]=R_{hw}-\|p_c(s[t],k_t)-O_{hw}\|;
\]
escape direction
\[
u_{\mathrm{esc}}=\frac{p_c-O_{hw}}{\|p_c-O_{hw}\|};
\]
and goal-alignment features \(a_0[t]\), \(a_1[t]\), and \(\Delta a[t]=a_0-a_1\), which encode target preference for **Switch**.

The full feature vector is assembled as
\[
x[t]=\mathrm{col}(x^{kin}[t],x^{ws}[t],x^{alg}[t],x^{ctx}[t],x^{loc}[t]).
\]
Empirically, this representation shift is consequential: on the canonicalization ablation, the **canonical frame** achieved **0.904 / 0.871** intent-recognition F1 on ID test / OOD-Reconfig, whereas the **world frame** yielded **0.889 / 0.614**, and **world + SE(2) aug** yielded **0.891 / 0.753** [2603.11077]. This suggests that the primary benefit is not simple augmentation, but the task-relative restructuring of geometry itself.

## 4. Task-aware Temporal Convolutional Network

TATIC feeds the causal temporal window
\[
X[t] \in \mathbb{R}^{T \times D}
\]
into a task-aware TCN. Inputs are standardized using training statistics, and independent dropout is applied to the localization subvector \([j[t]^\top\ s[t]]^\top\). A compact MLP encoder \(\phi_{\mathrm{enc}}\) maps each \(x[t]\) to an embedding \(z[t] \in \mathbb{R}^C\) with **channel width \(C=128\)**. Temporal processing uses **\(L=6\)** residual blocks with kernel size **\(k=3\)** and exponentially increasing dilations \(\{d_\ell\}\), followed by last-step pooling to produce the latent state \(h[t]\) [2603.11077].

The receptive field is
\[
R = 1 + (k-1)\sum_{\ell=1}^{L} d_\ell.
\]
For doubling dilations \(d_\ell=2^{\ell-1}\), \(\sum d_\ell = 63\), giving
\[
R = 127
\]
steps, which the paper states covers the full window. The network then branches into decoupled classification and regression heads,
\[
h^{cls}[t]=\phi_{cls}(h[t]), \qquad h^{reg}[t]=\phi_{reg}(h[t]).
\]
The classification branch outputs logits over the five operators and the two switch targets; the regression branch predicts \(\tilde d[t]\in\mathbb{R}^3\), \(\hat m[t]\), \(\hat \alpha_v[t]\), and \(\hat r_s[t]\), with domain clamping applied after prediction [2603.11077].

Training uses focal cross-entropy for operator classification, cross-entropy for target selection, masked cosine loss for Guide direction, masked MSE for scalar regressands, and homoscedastic uncertainty weighting across tasks:
\[
L = \sum_{u\in\mathcal{T}} \left(e^{-s_u}\,\bar L_u + s_u\right),
\]
where \(\mathcal{T}=\{\mathrm{op},\mathrm{dir},\mathrm{disp},\mathrm{spd},\mathrm{saf},\mathrm{tgt}\}\). The direction term is
\[
L_{\mathrm{dir}}[t]
=
\delta^{dir}[t]\left(
1 -
\frac{\hat d[t]^\top d_{loc}[t]}{\|\hat d[t]\|\,\|d_{loc}[t]\|}
\right).
\]
Optimization uses **AdamW with a cosine learning-rate schedule**, and the full model has approximately **0.63M trainable parameters** [2603.11077].

The architecture is explicitly causal and therefore designed for online deployment. In contrast to sequence models that depend on bidirectional context, TATIC constrains itself to the information available within the active contact segment.

## 5. Intent-conditioned adaptation and empirical validation

The output of TATIC is not merely a label; it is a control-relevant semantic state that drives distinct adaptation modes. For **Guide**, the executed path is
\[
\tilde x^{(r)}(\sigma) = x_d(\sigma) + d^{(r)}(\sigma),
\]
with semantic displacement
\[
\Delta x_{\mathrm{sem}}^{(r)} = m_{\max}\,\bar m[t_r]\,E[t_r]\hat d[t_r].
\]
The incremental displacement is injected over a bounded horizon using a bump function
\[
b(\xi)=64\xi^3(1-\xi)^3,
\]
ensuring endpoint preservation [2603.11077].

For **Yield**, the predicted safety radius is mapped into an inflated operation space via Minkowski sum,
\[
\tilde S_{hw}[t]=S_{hw}\oplus \mathbb{B}(\bar r_s[t]).
\]
For **Slow**, the reference velocity is scaled as
\[
v^{cmd}_{ref}[t]=\bar\alpha_v[t]\cdot v_{ref}[t],
\]
with a first-order low-pass filter enforcing acceleration limits. For **Switch**, the chosen mission goal is
\[
g^\star[t]=g_{\hat k[t]}[t],
\]
and the planner replans accordingly. **Stop** triggers a software-level hold at the current pose [2603.11077].

The evaluation dataset comprises a **7-DoF manipulator**, **500 in-distribution episodes across 100 trajectories**, and **250 out-of-distribution episodes across 50 reconfigured trajectories**. The ID split is **70/15/15** at trajectory level, with OOD-Reconfig reserved for OOD evaluation. On the test set, TATIC achieved **Macro-F1 = 0.904**, **ECE = 0.041**, and **95% bootstrap CI [0.887, 0.922]** for intent recognition. Per-class F1 scores were **Guide 0.862**, **Yield 0.879**, **Slow 0.926**, **Stop 0.901**, and **Switch 0.951** [2603.11077].

Regression performance was also reported. End-to-end results were **Guide direction cosine similarity 0.891**, **Guide magnitude RMSE 0.098**, **Slow scaling RMSE 0.079**, **Yield radius RMSE 0.108**, and **Switch target F1 0.921**. With ground-truth operator filtering, the corresponding values were **0.916**, **0.074**, **0.061**, **0.086**, and **0.938** [2603.11077].

Ablations isolate the role of feature groups. Using **kinematics only** gave **0.583 / 0.524 / 0.741 / 0.182** for operator Macro-F1, target F1, Guide direction cosine, and magnitude RMSE; adding **workspace** features improved this to **0.712 / 0.557 / 0.803 / 0.149**; adding **alignment** features gave **0.841 / 0.891 / 0.872 / 0.112**; and the **full TATIC** feature set reached **0.904 / 0.938 / 0.916 / 0.074** [2603.11077]. This indicates that semantic intent recognition in the framework is strongly dependent on task-relative geometric structure rather than kinematics alone.

Hardware validation was conducted in **collaborative desktop disassembly**, using a **7-DoF manipulator**, **side-mounted and wrist-mounted RGB-D cameras** for target coordinates, **torque-based contact estimation** for physical corrections, a **constraint-based task planner** from component connectivity and prerequisites, and **cuRobo** for nominal references. Reported examples include lateral pushes decoded as **Guide**, gentle pushes near a boundary decoded as **Yield**, taps decoded as **Slow**, directional corrections decoded as **Switch**, and assertive holds decoded as **Stop** [2603.11077]. The reported outcome was reliable semantic inference and responsive motion adaptation in hardware.

## 6. Limitations, failure modes, and nomenclature

The framework’s stated limitations are explicit. It uses a **predefined intent vocabulary** and does not yet model **personalization across operators**. Reported failure modes include **ambiguous corrections**, such as Guide versus Yield confusion for lateral pushes, and **contact estimation drift** under modeling errors or multi-contact scenarios. Planned extensions include adaptive personalization, expanded intent semantics, larger datasets, and multimodal fusion across vision, language, and force [2603.11077].

The acronym itself is not stable across adjacent literatures. In football analytics, **TATIC** is used as shorthand for **Tactical Style Imitation**, where the objective is to reproduce a team’s spatio-temporal tactical organization rather than infer intent from physical corrections; TacSIm operationalizes that usage as a benchmark grounded in broadcast footage and virtual-environment evaluation [2603.25199]. By contrast, the teamwork-intervention framework **ATTIC/TIC** is a different concept entirely: it learns a generative model of team behavior with BTIL and generates task-time interventions for coordination improvement, and the authors explicitly note that “TATIC” does not appear in that paper and is only a naming variant encountered elsewhere [2303.00413].

Within human-robot collaboration, however, TATIC is most precisely identified with the formulation in which torque residuals, contact geometry, task-frame canonicalization, and a causal TCN are combined to infer discrete operators and continuous task parameters from brief physical corrections. A plausible implication is that its main contribution lies in converting physical contact from a low-level compliance signal into a structured semantic control channel for online collaborative manipulation.

Source: https://www.emergentmind.com/topics/tatic