1 Introduction↩︎

Mobility impairments affect hundreds of millions of people worldwide, imposing significant burdens on both individuals and healthcare systems [1]–[3]. Powered lower-limb exoskeletons have emerged as a compelling assistive technology, able to support walking, stair climbing, and rehabilitation in clinical and daily settings [4]–[6].

Despite substantial hardware advances, the control of these devices remains a critical bottleneck. Traditional approaches rely on:

• Rule-based finite state machines—discrete mode graphs that are brittle on novel terrains or unexpected gait transitions [7], [8];

• Impedance/admittance controllers—manually tuned stiffness and damping parameters that not generalizable across users and different scenarios [9], [10];

• Offline-trained Weights—policies fixed at deployment time, unable to incorporate run-time feedback [11], [12].

Real-world mobility is inherently dynamic: terrain varies, users fatigue, gait patterns drift. Effective assistive systems must therefore adapt continuously to both environmental context and individual biomechanics.

We propose Olive, a learning framework that treats exoskeleton assistance as an online parameter-efficient adaptation problem. Drawing inspiration from low-rank adaptation in large-scale models [13], Olive confines real-time online updates to a compact low-rank subspace of the controller parameter space, so that to enable millisecond-level adaptation on wearable embedded hardware. Specifically, the frozen base controller \(W_0\) is initialised from our company’s internal motion foundation model pretrained on large-scale motion signals data encompassing electronic sensor streams (IMU, joint encoders, EMG), physical vibration signals, and diverse kinematic sequences—giving \(W_0\) a rich motion prior before any user-specific adaptation. Recent related work on efficient foundation models, scalable inference, long-context attention, multi-task adaptation, and experience-based planning [14]–[18] provides complementary context for edge-deployable online learning. Prior online approaches—including human-in-the-loop metabolic optimisation [6], [11] and graceful-degradation control [19]—address adaptation in isolation, without a unified treatment of stability, parameter efficiency, and personalisation. Olive addresses all three simultaneously within a single low-rank framework.

Contributions.We make the following contributions:

1. A low-rank adaptive controller built on a Pretrained-MM-initialised multimodal base policy, updated online via a reward-shaped policy gradient objective requiring no reference trajectories.

2. A gating mechanism that dynamically regulates the influence of the personalized residual based on user context.

3. A dynamic rank scheduler that selects rank \(r_t\) in real time from a candidate set based on estimated state complexity, balancing adaptation capacity against computational cost.

2 System Overview↩︎

Olive is deployed on our latest Wearable Exoskeleton, it is an ultra-lightweight bilateral hip-assist exoskeleton (\(\approx\)​2.4 kg, carbon-fiber + aerospace-grade aluminum alloy frame) with the specifications summarised in Table 1. Figure 1 shows the end-to-end pipeline.

5pt

Sensing layer. The system fuses four complementary motion modalities: (i) electronic signals — high-frequency IMU readings (acceleration, angular velocity at \(\sim\)​1000 Hz) and bilateral joint angle/velocity from encoders; (ii) surface EMG — bilateral muscle-activation signals providing a direct proxy for user intent and metabolic effort; (iii) physical vibration — ground-contact vibration sampled from the bilateral hip actuator force sensors, providing surface-texture cues unavailable from kinematics alone; (iv) context — an inferred activity-mode and terrain-class vector.

Intent recognition. A lightweight classifier maps the current sensor window to a discrete user intent \(u_t\!\in\!\{\textit{walk},\textit{climb},\textit{slope},\textit{uneven}\}\), running at \(<\)​2 ms on the embedded CPU.

Adaptive controller. The core OLIVE module (detailed in Section 3) maps the fused state \(s_t\) to assistive torque outputs \(a_t\in\mathbb{R}^m\) via a low-rank parameterisation updated online.

Embedded runtime. All inference and update steps run on a wearable ARM-based SoC, constrained to \({<}10\) ms round-trip latency to guarantee closed-loop safety.

Figure 2 illustrates Olive operating across four real-world scenarios encountered in daily use.

3 Method: Online Low-Rank Adaptive Control↩︎

3.1 Problem Formulation↩︎

At time step \(t\), the exoskeleton observes a multimodal motion state vector \[s_t= \bigl[\,x_t^{\mathrm{imu}},\; x_t^{\mathrm{joint}},\; x_t^{\mathrm{emg}},\; x_t^{\mathrm{vib}},\; x_t^{\mathrm{ctx}},\; h_{t-1}\,\bigr] \in \mathbb{R}^n, \label{eq:state}\tag{1}\] where \(x_t^{\mathrm{imu}}\) and \(x_t^{\mathrm{joint}}\) are high-frequency IMU and joint-encoder readings; \(x_t^{\mathrm{emg}}\!\in\!\mathbb{R}^{d_e}\) is a surface EMG signal capturing bilateral muscle activation, providing a direct proxy for user intent and metabolic effort; \(x_t^{\mathrm{vib}}\!\in\!\mathbb{R}^{d_b}\) captures ground-contact vibration signals from the actuator force sensors; \(x_t^{\mathrm{ctx}}\) encodes inferred terrain/activity context; and \(h_{t-1}\) is a summarised motion history. The system must produce an assistive action \(a_t\in \mathbb{R}^m\) (bilateral hip torques) via a policy \(\pi_{\Theta_t}\).

The key challenge is to continuously adapt \(\Theta_t\) during deployment while remaining computationally feasible on embedded hardware.

3.2 Low-Rank Adaptive Controller↩︎

We decompose the policy parameter matrix as: \[\Theta_t = W_0+ \Delta W_t, \qquad \Delta W_t= A_tB_t^\top, \label{eq:decomp}\tag{2}\] where \(W_0\!\in\!\mathbb{R}^{d\times k}\) is a frozen base controller initialised from our company’s motion foundation model pretrained on large-scale population-level data encompassing electronic sensor streams (IMU, encoders, EMG), physical vibration, and kinematic sequences—and held fixed throughout deployment; and \(A_t\!\in\!\mathbb{R}^{d\times r}\), \(B_t\!\in\!\mathbb{R}^{k\times r}\) are online-updatable low-rank factors with \(r\!\ll\!\min(d,k)\). The assistive action is thus \[a_t= \pi_{W_0+ A_tB_t^\top}(s_t). \label{eq:policy}\tag{3}\]

Computational savings. Full-matrix online updates would require maintaining and differentiating through \(\mathcal{O}(dk)\) parameters. By restricting updates to \(A_t\) and \(B_t\), the parameter count drops to \(\mathcal{O}(r(d{+}k))\), a reduction factor of \(\tfrac{dk}{r(d+k)} \gg 1\) for typical network widths \(d,k\) and small ranks \(r\in\{4,8,16\}\).

3.3 Gated Personalization↩︎

To prevent destabilising over-adaptation under uncertain or rapidly changing inputs, we introduce a scalar gating coefficient \[\alpha_t= \sigma\!\bigl(g(s_t, h_{t-1})\bigr) \in (0,1), \label{eq:gate}\tag{4}\] where \(g(\cdot)\) is a small two-layer network and \(\sigma\) is the sigmoid function. The gated policy becomes: \[\boxed{a_t= \pi_{\;W_0+ \alpha_t\,A_tB_t^\top}(s_t).} \label{eq:gated}\tag{5}\] When \(\alpha_t\!\approx\!0\) (stable, familiar context), the system falls back to the robust base controller; when \(\alpha_t\!\approx\!1\) (detected state shift or user fatigue), the personalized component is fully activated.

3.4 Dynamic Rank Scheduling↩︎

Flat walking is geometrically simple and well-covered by a low-rank update, whereas complex transitions—stairs, uneven terrain—demand richer representational capacity. Fixing \(r\) globally either wastes computation on easy contexts or starves capacity on hard ones. We therefore introduce a lightweight complexity estimator \[c_t = \sigma\!\bigl(\psi(s_t, h_{t-1})\bigr) \in (0,1), \label{eq:complexity}\tag{6}\] where \(\psi(\cdot)\) is a single-hidden-layer MLP that shares its first-layer weights with the gating network \(g(\cdot)\) to minimise parameter overhead. The effective rank is then discretised from a candidate set \(\mathcal{R}=\{r_{\min},\ldots,r_{\max}\}\): \[r_t = \mathrm{clip}\!\Bigl( \bigl\lfloor c_t \cdot (|\mathcal{R}|)\bigr\rfloor + r_{\min},\; r_{\min},\; r_{\max}\Bigr). \label{eq:rank}\tag{7}\] We maintain low-rank factors at maximum capacity \(A_t\!\in\!\mathbb{R}^{d\times r_{\max}}\), \(B_t\!\in\!\mathbb{R}^{k\times r_{\max}}\) and truncate to the leading \(r_t\) columns at inference: \[\Delta W_t= A_t[\text{:,}\,{:}r_t]\;B_t[\text{:,}\,{:}r_t]^{\!\top}, \label{eq:dyn95delta}\tag{8}\] so that the gated policy of Eq. 5 becomes \(a_t= \pi_{W_0+ \alpha_t\,\Delta W_t}(s_t)\). In practice we set \(r_{\min}{=}4\), \(r_{\max}{=}16\), giving a 4\(\times\) range in online parameter count: simple flat walking converges to \(r_t{=}4\) (\(4{\times}(d{+}k)\) multiply-adds), while novel uneven terrain expands to \(r_t{=}16\) on demand. This co-design of the gating and rank signals with shared representations keeps the added overhead under 0.3 ms per step on the embedded SoC.

3.5 Online Incremental Update via Reward-Shaped Policy Gradient↩︎

We design OLIVE as an online adaptation method based on policy gradient method in the low-rank subspace. At each step \(t\) the exoskeleton executes \(a_t\), observes on-body sensory feedback, and receives a shaped reward \[r_t = w_1\,\Delta\overline{\mathrm{EMG}}_t + w_2\,(1-\mathcal{E}_t) + w_3\,(1-\|\phi(s_t,a_t)\|_2), \label{eq:reward}\tag{9}\] where \(\Delta\overline{\mathrm{EMG}}_t = \overline{\mathrm{EMG}}_{t-1} - \overline{\mathrm{EMG}}_t\) is the inter-step decrease in mean bilateral muscle activation (positive when the exoskeleton reduces effort, negative when it increases it); \(\mathcal{E}_t \in [0,1]\) is a normalized proxy signal of metabolic effort from IMU variance and contralateral load asymmetry; and \(\phi(s_t,a_t) \in [0,1]\) measures normalised CoM deviation and bilateral step asymmetry. All three terms are normalised to \([0,1]\) before weighting, so \(r_t\) is dimensionless and bounded. Because \(r_t\) is computed entirely from on-body sensors, Olive does not require offline trajectories.

The low-rank factors \(A_t,B_t\) are updated by minimising a regularised negative-reward objective: \[\mathcal{L}_t = -\lambda_1 r_t + \lambda_2\mathcal{L}_{\mathrm{smooth}} + \lambda_3\mathcal{L}_{\mathrm{stab}}, \label{eq:loss}\tag{10}\] where the regularisers penalise torque discontinuities and postural instability: \[\begin{align} \mathcal{L}_{\mathrm{smooth}} &= \|a_t- a_{t-1}\|_2^2, \tag{11} \\ \mathcal{L}_{\mathrm{stab}} &= \|\phi(s_t,a_t)\|_2^2. \tag{12} \end{align}\] Minimising \(\mathcal{L}_t\) via gradient descent on \(A_t,B_t\) is equivalent to a one-step proximal policy gradient update [20] within the low-rank manifold, with the regularisers acting as a trust-region constraint that prevents destabilising jumps.

The incremental gradient update step is: \[A_{t+1} = A_t- \eta\,\nabla_{A_t}\mathcal{L}_t, \qquad B_{t+1} = B_t- \eta\,\nabla_{B_t}\mathcal{L}_t, \label{eq:update}\tag{13}\] with \(W_0\) held fixed throughout deployment. Stability.Because \(W_0\) is frozen, the full policy perturbation is bounded by \(\|\alpha_t\,\Delta W_t\|_F \leq \|A_t\|_F\|B_t\|_F\); the gating \(\alpha_t\!\in\!(0,1)\) and step size \(\eta\) together ensure the residual norm remains bounded under gradient descent, providing a Lyapunov-like stability guarantee [21].

4 Preliminary Evaluation↩︎

4.1 Setup↩︎

We evaluate Olive on the excoskeleton with six healthy participants (age 24–38, 3F/3M). Each participant completed three sessions (\(\sim\)​5 000 walking steps per session) covering four terrain conditions in sequence: flat walking (100 m), stair ascent/descent (5 flights), sloped terrain (\(8^{\circ}\)), and uneven cobblestone. All four sensing modalities were active: the onboard 1000 Hz IMU and joint encoders, bilateral surface EMG (8-channel Delsys Trigno), and actuator force sensors for vibration. We compare against three baselines: Static—a fixed gait cycle controller; Rule-Based—a finite-state machine with terrain detection; Fixed-NN—a neural policy trained offline on population data and initialised from Pretrained-MM(same backbone as Olive, without online updates).

Metrics include gait smoothness (inverse of acceleration jerk norm, normalised), effort reduction (a metabolic proxy relative to unassisted walking), and motion stability (inverse CoM variance, normalised). All metrics are min-max normalised to \([0,1]\).

4.2 Results↩︎

Figure 4a presents the full evaluation. Olive achieves a smoothness score of 0.93 with a +13% improvement over the strongest baseline. Effort reduction is 0.67 vs. (+22%), and motion stability is 0.91 vs. (+15%). The learning curve (Figure 4b) tracks the normalised user-effort score . Olive converges to an effort score of \(0.33\) (effort reduction \(0.67\)) within approximately 1 800 steps—40% below Fixed-NN’s final score of \(0.55\). Static and Rule-Based controllers show negligible improvement over time. Terrain generalisation results (Figure 4c) confirm that Olive retains superior smoothness on stairs, slopes, and uneven surfaces. The low standard deviation of Olive across all conditions indicates robust adaptation rather than overfitting to flat terrain. The radar chart (Figure 4d) further shows that Olive comprehensively dominates on all five dimensions: smoothness, effort reduction, stability, adaptation speed, and terrain generalisation.

Ablation. Table 3 isolates each component’s contribution. Removing weights initialisation causes the largest drop (\(-\)​0.11 smoothness, \(-\)​0.14 effort reduction). Removing gating disproportionately degrades stability (\(-\)​0.07 vs. \(-\)​0.05 smoothness), which consistent with the gating role in suppressing residual overshoot on terrain transitions. Removing dynamic rank scheduling has minimal effect on quality (\(-\)​0.01 smoothness) but increases latency by 0.8 ms, so it validats the online ranking efficiency benefit.

Efficiency. With \(d{=}k{=}128\) and \(r_{\max}{=}16\), the worst-case update parameter count is \(16{\times}(128{+}128) = 4{,}096\) vs.\(128{\times}128 = 16{,}384\)—a \(4\times\) reduction even at full rank. During typical flat-walking sessions, the dynamic rank scheduler converges to \(r_t{=}4\) for over 70% of steps, yielding an effective \(16\times\) reduction in those contexts and reducing mean per-step update cost to 1.3 ms. End-to-end inference-and-update latency is 7.4 ms on the embedded SoC, comfortably within the 10 ms safety budget, with rank scheduling adding only \({\approx}0.3\) ms overhead.

4.5pt

3.5pt

5 Conclusion↩︎

Olive validates that we can perform online weight updates in a low-rank subspace as a balanced approach to stability and personalization. OLIVE uses a reward-shaped policy gradient to eliminate reference-trajectory dependence and it extends the human-in-the-loop metabolic optimisation paradigm [6], [11] to a fully online low-rank update setting. Furthermore, we improve stability during terrain transitions via gating, and reduce computation on flat terrain (\(r_t{=}4\), \(>\)​70% of steps) through dynamic rank scheduling, while expanding capacity on complex surfaces (\(r_t{=}16\)). Together, these form OLIVE as a hardware efficient method of whether and how much to update weights online for wearable devices, improving both stability and personalization.

References↩︎