Organizations: University College London, London, United Kingdom · Institute of Automation, Chinese Academy of Sciences (CASIA), Beijing, China · University of Chinese Academy of Sciences (UCAS), Beijing, China · Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China
Embodied continual learning requires controllers to infer changing situations, retain prior behavior and reuse motor skills without task labels or known boundaries. The challenge is deciding when to recover earlier behavior and which skills can support new learning. We introduce Situation-guided Primitive Integration for Recurrent Embodiment (SPIRE), a recurrent-spiking system that infers latent contexts from action-outcome evidence and learns context-specific motor actors online. Mature actors become protected sources that combine with a target-context residual to drive recurrent-spiking motor populations. Across coordinate, three-joint and hidden-force control, SPIRE maintained control, autonomously inferred situations, retained prior behavior and accelerated acquisition of an unseen source composition, reducing mean interactions to criterion by approximately 40% relative to matched fresh controllers. Rewinding compatible source actors impaired early acquisition in every matched intervention, demonstrating causal reuse. Against seven recurrent continual-learning and modular-control comparators, SPIRE achieved 2.4 times the mean prior-learning advantage of the strongest eligible task-free alternative selected for each paired stream. These results demonstrate autonomous recovery and causal reuse of motor skills for new learning in continuous embodied control.
Figures & tables
Fig. 1: SPIRE learning architecture. From left to right: situation inference, primitive composition and recurrent-spiking action output. Colored dashed arrows indicate selection of source and residual actors, while the solid xt arrow carries their current input. Paired event rows denote positive and negative population channels. Environment feedback supports context inference and local learning. The context-belief map and control environments are schematic.
Fig. 2: Three embodied control forms and the continuous task sequence. (a) Coordinate regulation uses axial source targets A and B, an opposing target X and the unseen diagonal composition C. (b) Joint control uses one three-link body and base. Gray poses illustrate A, B and X, and the navy pose illustrates C; rings mark ideal target endpoints, with each arm shown near its target. (c) Hidden-force compensation stabilizes the body at the origin against the spatially uniform C disturbance (+x,+y) . (d) The sequence is illustrated with hidden forces while the origin target remains fixed: source skills are learned in A and B, X introduces interference, C combines the A and B disturbances to test reuse, and return-A tests retention. Target rings, force arrows and compensation arrows follow the shared legend. Paths and compensation directions are schematic. Controller memory persists throughout the sequence, and phase labels are not supplied to SPIRE.
Fig. 3: Context inference, control, retention and source reuse. Each form has eight replicates per method; points show individual runs and short bars show means. (a) Mean occupancy of every observed latent context across 24 SPIRE runs. Each phase includes all 640 assignments per run, and each column sums to 100%; return-A recovers C0. (b) Late phase scores across A–B–X–C–A, with form means connected. (c) Immediate and late return-A degradation; negative values indicate improvement and dashed lines mark the retention limits. (d) Early-C score advantage over a same-initialization fresh controller. (e) Transactions to the C criterion; lines pair learned and fresh runs, with mean reductions annotated. (f) Early-C score loss after rewinding A+B, A, B or X source actors. Green marks show reuse thresholds; shading marks the X tolerance of ±0.03 . (g–i) Prior-learning advantages across methods in coordinate, joint and force control. Oracle EWC and SI receive phase-start information.
Fig. 4: Protected source contributions during C acquisition. Each form contains eight matched learned/fresh pairs. (a) Selected high-benefit examples over the first 128 C transactions: learned-history controller (blue) and same-initialization fresh controller (coral). Saturation indicates elapsed time; dots mark endpoints. Gray and blue arm poses show the initial and learned-history final three-joint pose. (b) Protected source A/B and active-context actor R contributions over 192 C transactions, before clipping, exploration and recurrent event decoding; R becomes the target residual once composition is selected. Signed confirmation probes contribute no actor activity. Eight-transaction moving means are averaged across replicates. Each form uses a symmetric color scale at the 98th percentile of absolute mean contributions: blue is positive, coral negative and white zero. The upper trace shows mean ∥uA+uB∥2 with a ±1 sample-SD envelope. (c) Learned/fresh scores: 16-transaction moving means with ±1 sample-SD envelopes, clipped to the score range [0,1]. Annotations give unsmoothed mean paired gaps and positive-pair counts.
Method
Coordinate
Three-joint
Hidden force
Immediate
Late
Immediate
Late
Immediate
Late
Oracle EWC
13.1±7.1
−5.2±2.2
10.4±5.6
−4.7±2.4
2.4±2.7
−5.3±3.1
Oracle SI
9.3±3.4
−6.1±1.6
12.7±7.4
−4.2±1.6
3.8±4.4
−5.2±2.4
Naive recurrent ANN
8.5±7.1
−6.5±2.0
9.2±3.6
−4.0±1.8
6.4±4.7
−4.7±1.5
Task-free EWC
9.2±4.5
−1.0±4.1
14.0±7.2
−3.0±2.0
3.1±2.2
−4.2±1.5
Task-free SI
5.7±7.1
−7.4±2.9
7.4±3.4
−4.9±1.4
5.5±3.6
−4.4±1.7
TABLE I: Return-A degradation across methods (%). Mean ± sample SD of score loss from initial-A final 256 transactions to return-A first 128 (immediate) or final 256 (late), across eight runs per method and form. Negative values indicate improvement. Bold values mark the lowest mean among task-free methods in each column. The two oracle references above the rule receive phase-start information.
Coordinate
Three-joint
Hidden force
Variant
Return A
Early C
Return A
Early C
Return A
Early C
Shared context
0.577±0.222
0.169±0.115
0.660±0.163
0.395±0.202
0.831±0.025
0.494±0.268
No composition
0.864±0.023
0.553±0.051
0.861±0.032
0.561±0.101
0.825±0.028
0.484±0.062
No recurrence
0.855±0.026
0.677±0.025
0.856±0.027
0.695±0.017
0.831±0.027
0.712±0.030
SPIRE (full)
0.864±0.023
0.678±0.026
0.861±0.032
0.701±0.018
0.825±0.028
0.714±0.028
TABLE II: Behavioral effects of module ablations. Mean and sample SD (below) over eight paired runs per form and variant. Return A uses the first 128 return transactions, and Early C the first 192 C transactions. Both scores range from 0 to 1; higher is better. Bold means mark column maxima, including ties.
Intervention
Coordinate
Joint
Hidden force
Output-event suppression
0.2378±0.0137
0.2133±0.0101
0.2507±0.0148
Recurrence removal
0.0132±0.0026
0.0129±0.0014
0.0129±0.0035
Private-state reset
0.0124±0.0045
0.0099±0.0008
0.0093±0.0030
TABLE III: Effects of recurrent motor interventions. Mean absolute changes in normalized action commands relative to matched intact execution, reported as means ± sample SD across eight SPIRE runs per form. Each run contributes one 32-step intervention comparison.
Traditional robot programming is notoriously challenging: it requires orchestrating multimodal perception, managing physical contact dynamics, and handling diverse environment configurations and execution failures. We introduce ASPIRE (Agentic Skill Programming through Iterative Robot Exploration), a continual learning system that autonomously writes and refines robot control programs in a code-as-policy paradigm while compounding experience into a reusable skill library. ASPIRE operates in an open-ended learning loop, consisting of three components: (1) a closed-loop robot execution engine that exposes fine-grained multimodal traces, enabling the agent to autonomously diagnose failures, synthesize repairs, and validate outcomes; (2) a continually expanding skill library that distills validated fixes into reusable, transferable robotic knowledge; and (3) an evolutionary search procedure that generates diverse programs, systematically debugging them to explore beyond single-trajectory refinement. ASPIRE surpasses prior methods by up to 77% on manipulation tasks under perturbation, 72% on Robosuite's bimanual handover task, and up to 32% on long-horizon household tasks. The accumulated skill library further enables zero-shot generalization: on representative unseen long-horizon tasks, ASPIRE achieves 31% success, outperforming the 4% success rate of prior methods. Performance scales with the capabilities of frontier models. With Claude Opus 5.5, ASPIRE nearly saturates these benchmarks, achieving 99% average success across all 80 LIBERO-Pro tasks and 100% success on representative Robosuite and BEHAVIOR-1K tasks. Beyond simulation, skills discovered in simulation reduce real-world robot programming effort across embodiments and robot APIs, while skills acquired through online interaction with physical robots enable ASPIRE to solve a Rubik's Cube with progressively improving efficiency.
Embodied Continual Learning (ECL) aims to enable robots to continually acquire new manipulation tasks while retaining previously learned behaviors under closed-loop control. In ECL, feature drift can propagate through sequential decision-making under closed-loop control, turning representation changes into compounding behavioral deviations on previously learned tasks. A key challenge in ECL lies in structured skill reuse across continually evolving tasks, since existing methods primarily focus on skill learning without explicitly organizing them for coherent task execution. To address this issue, we propose SCE, a Skill-Compositional Experts framework for ECL. SCE builds a skill base via Compositional Skill Grounding (CSG), which decomposes task demonstrations into reusable skills. Based on this, Dual Execution-and-Transition Experts (DETE) enable new task learning through skill composition, where one branch ensures skill execution and the other supports transitions between skills for coherent behavior. Experiments on LIBERO benchmarks and real-world manipulation tasks show that SCE improves retention and overall task performance. Further feature drift analyses and ablation studies verify the effectiveness of our method. Project website: https://eqcy.github.io/sce/.
Shuaike Zhang, Shaokun Wang, Haoyu Tang +2
Harbin Institute of Technology, Shenzhen · Shandong University · Shenzhen Loop Area Institute
Humanoid whole-body controllers can now track a diverse set of dynamic motions, but they are typically trained offline and then frozen, so teaching such a controller a new skill tends to erode the skills it already mastered. We study continual learning for humanoid whole-body motion, where a single controller must acquire skills from a sequential task stream without revisiting past data. We introduce Similarity-guided LoRA-PNN, a progressive neural network (PNN) policy that prevents catastrophic forgetting by construction while reusing knowledge across skills through lightweight low-rank adaptation. A two-level motion-similarity measure, built from dynamic time warping aggregated by optimal transport, decides which prior skill to build on and how much new capacity to allocate, yielding strong forward transfer and large efficiency gains. Across six sequentially learned skill categories, our similarity-guided LoRA policy attains the best forward transfer (0.125 vs. 0.079) and the highest average accuracy among all methods, while saving up to 94.5% of trainable parameters and 40.8% of training time. The resulting controller reaches 96.13% sim-to-sim transfer and is deployed on a physical Unitree G1. Our code is available at https://anonymous.4open.science/r/continual-humanoid-learning-35D3.