Restoring Precision Grasping and Tactile Sensation via a Closed Loop BCI & Reinforcement Learning- controlled Active Orthosis
DOI: Pending DOI
Synopsis
Recovering hand function, both movement and touch, is among the most pressing goals in modern neurotechnology for individuals living with paralysis. We have previously shown that a bidirectional neural bypass (BNB), which links the brain directly to the body, can produce meaningful gains in motor and sensory ability[1]. The standard BNB approach relies on neuromuscular electrical stimulation to drive movement; however, this modality is ineffective in a subset of individuals, including those with lower motor neuron damage. To overcome this limitation, we augmented the BNB with a lightweight, custom-fabricated active orthosis (AO) capable of mechanically actuating the fingers and providing real-time tactile sensory feedback. A participant with complete C4/C5 tetraplegia and lower motor neuron dysfunction was fitted with a 3D-printed, low-profile splint housing, a servo-driven tendon-like actuator and an embedded force sensor. Cortical signals captured by intracortical microelectrode Utah arrays were decoded online through deep recurrent neural networks (LSTMs) to identify the participant's intended grasp action. Once a grasp was initiated, a deep Q-network (DQN) reinforcement learning agent continuously adjusted actuator position to keep contact force within safe limits. Simultaneously, intracortical micro stimulation (ICMS) of primary somatosensory cortex conveyed force-related tactile percepts back to the participant. With this closed-loop system, the participant grasped, lifted, and released everyday objects at over 80% accuracy, - the first time he had done so independently since his injury.
Background
Each year, roughly 18,000 people in the United States sustain a spinal cord injury (SCI), and a substantial proportion of these injuries compromise upper-limb function, severely limiting personal independence. Brain-computer interface (BCI) research has established that cortical motor signals can be harnessed to drive a range of effectors, from robotic arms to the user's own limb[2]. Neuromuscular electrical stimulation is one widely explored approach for restoring movement, yet it remains unsuitable for every individual, particularly those with lower motor neuron injury. Alternative assistive devices, such as powered orthoses and exoskeletons, can partially compensate for lost motor function, but current designs tend to be bulky, require lengthy donning procedures, and operate without sensory feedback. The absence of real-time tactile information forces users to rely on vision, increasing cognitive demand and making delicate, everyday tasks impractical[3].
These device limitations underscored the need for a system that tightly couples an assistive device with a neural interface. Here we present a bidirectional neural bypass paired with a reinforcement-learning-controlled active orthosis that allows a person with complete C4/C5 tetraplegia and lower motor neuron dysfunction to perform precision grasps of fragile objects, prerequisite for activities of daily living (ADLs) such as eating and drinking, guided solely by decoded cortical intention and cortically delivered sensory feedback.
Methods
The study is registered at ClinicalTrials.gov (NCT03680872) and was conducted under an Investigational Device Exemption (IDE G170200) issued by the US Food and Drug Administration (FDA). The protocol was approved by the Northwell Health Institutional Review Board (IRB# 17-0840) and was in accordance with the Declaration of Helsinki.
System Overview & Nested Control Architecture- Achieving functional grasp in complete tetraplegia demands closed-loop coordination of movement and sensation. Our BNB system pairs an RL-controlled active orthosis with intracortical recording and stimulation to restore both volitional hand movement and touch. Intracortical microelectrode arrays in primary motor cortex (M1) capture neural activity, which is decoded online by an LSTM to classify intended actions (Open, Close, Reach, Rest). A decoded grasp command engages a servo-driven tendon mechanism on the AO, flexing the fingers into a power grasp. An embedded force sensor then continuously monitors contact force and streams it to a DQN RL agent, which adjusts servo position in real time to hold force within safe bounds and averting both slippage and object damage. This nested architecture mirrors primate sensorimotor hierarchy[4]: the LSTM resolves intent while the RL agent handles force regulation, reducing cognitive load on the user. To complete the loop, force exceeding an object-specific threshold triggers ICMS in S1 via separately implanted arrays, evoking a tactile percept in the index finger that confirms a secure grasp.
Participant & Neural Interface- The participant (male, 40s) sustained a complete C4 sensory/C5 motor SCI in 2020 from a diving accident. In 2023, five 10×10, 1.5 mm platinum Utah arrays (Blackrock Neurotech) were implanted in left-hemisphere M1 and S1[5]. Transcutaneous spinal cord stimulation (C5, 60 min; C7, 30 min; 300 mA) paired with 10 weeks of activity-based training partially restored proximal arm function[6], but the hand remained paralyzed, necessitating the AO for grasp restoration.
Active Orthosis Design & Fabrication- Effective grasping in ADLs requires continuous force modulation tuned to object stiffness, weight, and texture[7]. Because power grasp is biomechanically dominated by coordinated finger flexion against the thumb, this multi-joint action can be represented as a low-dimensional output, a simplification we exploited in designing the AO. The participant's hand and forearm were 3D-scanned, and a custom orthosis skeleton was digitally designed (Spentys, Belgium), then printed in biocompatible Nylon 12 (Formlabs Fuse 1). After fit testing for anatomical conformity and comfort, the shell was modified to add securing straps, a forearm-mounted 14 kg-pull servo (Injora), and a thumb-mounted load-cell sensor (Honeywell FMA 5N). The servo drives a string-and-magnet tendon system that flexes the fingers and the magnets disengage above 8 lbs to safeguard the user. The force sensor sits between the thumb knuckle and the rigid nylon body, ensuring consistent force transfer across donning sessions.
Motor Intent Decoding- M1 signals were decoded online by an LSTM network[8] into four classes (Open, Close, Reach, and Rest) which triggered servo actuation. The servo advanced at a fixed velocity upon each decoded movement of close hand. This scheme sufficed rigid objects but lacked the graded control needed for fragile items, motivating the addition of the RL force controller.
Deep Q-Network Reinforcement Learning Agent- Drawing on the hierarchical sensorimotor loops of primates[4], we let decoded intent initiate the servo while a dual DQN RL agent, trained offline in MATLAB Simulink, regulated force in real time. The agent received continuous force at 10 Hz and output servo commands to keep grasp force within empirically derived bounds. Stiffness and damping of the AO-hand system were characterized through repeated eggshell grasps to define acceptable force limits. The reward function incentivized maximum servo speed below the minimum force threshold and rapid deceleration above the maximum, enabling robust on/off force control within just 8 training episodes (2 trials/episode). White noise was added to simulated sensor data to ensure robustness.
Sensory Feedback via ICMS- ICMS was delivered through a Cerestim R96 stimulator (Blackrock Neurotech). Percept mapping across the S1 arrays identified electrodes that evoked sensation in the index finger. When the AO force sensor exceeded an object-specific threshold, those electrodes were stimulated at 100 µA, producing a clear fingertip percept that gave the participant confidence in his grasp.
Experimental Setup- The participant's view of his hand was occluded by a cardboard screen while he grasped hollowed eggshells. A real-time force trace on a monitor indicated the target range; a trial was scored successful only if the force stayed within bounds. Across 182 trials (RL-on and RL-off conditions), we assessed force control accuracy. In a separate block, S1-ICMS feedback was toggled: the participant was instructed to lift the eggshell only when confident the grasp was secure. Success rates were compared with unpaired two-tailed t-tests (α = 0.05).
Results
In the absence of the RL agent, grasp force was highly variable, frequently exceeding the predefined safe limit. Engaging the deep RL agent markedly improved performance: the participant achieved an 87% success rate with the agent active compared to 27% without it (p=0.012, unpaired t-test), confirming that real-time RL-based servo regulation substantially enhanced force control precision. When S1-ICMS somatosensory feedback was introduced, task accuracy rose further, reaching 100% compared to chance-level performance in the no-feedback condition (p<0.01). Thus, demonstrating that cortically delivered tactile information meaningfully informed the participant's grasp decisions.
Beyond controlled laboratory trials, the fully integrated BCI–AO system enabled the participant to complete functional activities of daily living. In one demonstration, he grasped a beverage can, lifted it, and took a sip; in another he held a soft food item and brought it to his mouth to eat. Throughout both tasks, the AO maintained grasp force within safe bounds, and the participant reported that the ICMS-evoked tactile percept gave him the confidence needed to lift and manipulate each object. These demonstrations marked the first time the participant had independently eaten and drunk using his own hand in the three years since his injury.
Discussion
Spinal cord injury can produce paralysis that profoundly diminishes independence and quality of life. For affected individuals, recovery of hand function consistently ranks among the highest rehabilitation priorities[9], yet the options available to those with complete injuries remain extremely constrained. To meet this critical clinical need, we designed and implemented a bidirectional neural bypass that unites an intracortical BCI with an RL-controlled active orthosis, restoring volitional hand movement and tactile sensation in parallel and enabling a person with tetraplegia to carry out every day functional tasks.
A persistent obstacle in iBCI-mediated hand control has been the precise regulation of grasp force demanded by real-world activities, handling fragile items, soft foods, and beverage containers. By introducing a novel deep reinforcement learning framework, we demonstrated a proof-of-concept system that rapidly stabilizes contact force, allowing reliable interaction with delicate objects. The nested control architecture of pairing a stable LSTM decoder for motor intent with a DQN agent for fine force regulation draws on the distributed, hierarchical organization of primate sensorimotor circuits[8]. Notably, the participant was able to perform grasping tasks while simultaneously holding conversation, pointing to a low cognitive burden and suggesting strong potential for deployment outside the laboratory. Although the system was validated in a single participant, it is readily transferable: a new user requires only a brief hand scan to produce a custom-fit orthosis, while the LSTM decoder and RL agent can be reused with minimal retraining, lowering the barrier to adoption across a broader SCI population. Looking ahead, the BNB platform may also serve as a vehicle for therapeutic stimulation in other individuals with SCI, pairing cortical engagement with collaborative motor tasks mediated by the RL-based AO to promote neuroplasticity and long-term functional gains[10].
Disclosures
CEB has financial interests in Neuvotion, Inc, a medical device company developing neurotechnologies for restoring function after stroke or spinal cord injury, and Sanguistat, LLC, a company developing neurostimulation technologies to reduce blood loss; he also has multiple patents in the field of neuroprosthetics and related fields. The remaining authors declare no competing interests.
References
Bouton, C. E. et al. (2016). Restoring cortical control of functional movement in a human with quadriplegia. Nature 533, 247–250.
Chandrasekaran, S. et al. (2021). Historical perspectives, challenges, and future directions of implantable brain-computer interfaces for sensorimotor applications. Bioelectron Med 7, 14. doi: 10.1186/s42234-021-00076-6.
Godfrey, S. B., Bianchi, M., Bicchi, A. & Santello, M. (2016). Influence of force feedback on grasp force modulation in prosthetic applications: a preliminary study. Annu Int Conf IEEE Eng Med Biol Soc 2016, 5439–5442. doi: 10.1109/EMBC.2016.7591957.
Oya, T., Takei, T. & Seki, K. (2020). Distinct sensorimotor feedback loops for dynamic and static control of primate precision grip. Commun Biol 3, 156.
Chandrasekaran, S. et al. (2021). Evoking highly focal percepts in the fingertips through targeted stimulation of sulcal regions of the brain for sensory restoration. Brain Stimulation 14, 1184–1196.
Chandrasekaran, S. et al. (2023). Case study: persistent recovery of hand movement and tactile sensation in peripheral nerve injury using targeted transcutaneous spinal cord stimulation. Front Neurosci 17, 1210544. doi: 10.3389/fnins.2023.1210544.
Nowak, D. A. & Hermsdörfer, J. (2005). Grip force behavior during object manipulation in neurological disorders: toward an objective evaluation of manual performance deficits. Mov Disord 20, 11–25. doi: 10.1002/mds.20299.
Chandrasekaran, S., Wandelt, S.K., Jangam, A. et al. A neuroprosthesis for restoring hand movement and sensation in a person with complete tetraplegia. Nat Med 32, 2591–2601 (2026). https://doi.org/10.1038/s41591-026-04498-0
Anderson, K. D. (2004). Targeting Recovery: Priorities of the Spinal Cord-Injured Population. Journal of Neurotrauma 21, 1371–1383.
Cater, E. et al. (2025). Cortically Interfaced Human Avatar Enables Remote Volitional Grasp and Shared Discriminative Touch. 10.1101/2025.09.21.25336267.