Continuous Lower-Limb Joint-Angle Estimation from sEMG and IMU: The Conditional Value of Cross-Activity and Cross-Subject Data under Update-Matched Controls

TL;DR An offline study that estimates hip, knee and ankle angles from leg sEMG and an IMU, and measures which extra training data actually help a new user against a control matched in training updates. Adding the user's own other activities raises R² from 0.722 to 0.788, of which +0.048 is information rather than extra training.
What they do
Exoskeletons driven by the wearer’s intent need a continuous, causal estimate of joint angles, yet a new user brings little calibration data. This study asks which extra data are worth adding: the user’s recordings of other activities, or other users’ recordings of the same activity. It uses a public dataset of ten healthy adults (64-channel high-density sEMG below the right knee, a tibia IMU, six locomotion activities with three gait-cycle records each). Everything is evaluated offline; no exoskeleton is used.
How it works
- Frozen decoder. A strictly causal network with 76,534 parameters reads 0.5 s of time-domain sEMG and IMU features through residual causal convolutions and a GRU, and outputs hip, knee and ankle angles. It is fixed before any data experiment.
- Update-matched replay. More data also means more optimizer updates. Each added source is compared with training on the user’s own two target records for the same number of updates, which separates information from training budget.
- Held-out trials. One trial of the target activity is always held out. Effects are averaged per participant, with bootstrap CIs and exact sign-flip tests.
- Other users’ data serve as pretraining before fine-tuning, with or without the user’s other activities.
Results
- Own other activities: R² 0.722 → 0.788 and RMSE 6.68° → 5.63°. Replay alone reaches 0.740, so 27% of the naive gain is budget. Stair descent is a reliably harmful donor activity (−0.060 R²), sit-stand-walk a beneficial one (+0.036).
- Other users’ data add +0.053 R² when the user has only the target activity, and +0.028 once own other activities are included. Other users, then own activities, then the target reaches R² 0.810.
- In a separate analysis on six participants, donor data add +0.47 R² with no same-activity record, nothing with one, and −0.036 with two.
Limitations
Offline only: no exoskeleton, closed-loop control or online latency is tested. Ten participants and six activities from one dataset; the full-calibration analysis uses six participants and a different model. The activity label is assumed known at inference.
Figures



Abstract
Intent-driven exoskeleton control needs a continuous, causal estimate of the wearer's joint kinematics and an answer to a second question: whose data should train the estimator for a new user? Borrowing the user's recordings of other activities, or other users' recordings of the same activity, is usually evaluated as "more data" versus "less data," which confounds the information a source carries with the optimizer updates it adds and hides negative transfer behind averages. We separate these effects for hip, knee, and ankle angle estimation from high-density sEMG and an IMU across six locomotion activities on all ten participants of a public dataset. A compact residual causal TCN-GRU decoder is frozen, and an update-matched replay control prices every added source against the same number of updates on the user's own data. A user's own other activities raise $R^2$ from 0.722 to 0.788, of which +0.048 is information beyond replay; stair descent is a reliably harmful donor and sit-stand-walk a reliably beneficial one. Other users' same-activity data help while the user has only the target activity ($G_\mathrm{T} = +0.053$) and half as much once own other-activity data exist ($G_\mathrm{W} = +0.028$); at full calibration, donor value collapses after one to two target records. Per-target gains from the two sources correlate, and leave-one-participant-out gates recover most of the cross-subject value only at low calibration. Data value is hierarchical, state-dependent, and directional; the bottleneck is deciding, without test labels, when to transfer.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
