Intention-to-Action (I2A)
This is the core chapter of the entire BCI topic. It weaves together all the decoding techniques from earlier chapters (04, 05) and all the downstream applications (07, 08, 09) into a single complete pipeline:
neural signals → intent extraction → action planning (LLM / POMDP) → execution → sensory feedback → neural signals ...
Traditional BCIs stopped at kinematic decoding (decoding joint angles, velocities, forces). Modern BCIs couple intent decoding with LLM / RL planners, so the user only has to think "I want a drink of water" rather than "bend the elbow 30°, rotate the wrist 45°." This is exactly what sets the "intention-to-action" paradigm apart from the previous generation of BCIs.
Why this chapter matters. The I2A paradigm upgrades BCI from a "control interface" to an "agent interface" — the user no longer operates the end-effector but expresses a high-level goal, which an LLM / POMDP translates into a low-level action sequence. This is also the bridge that lets Chapter 10 connect BCI to embodied and human-like intelligence: when intent itself is the input to the planner, BCI becomes the interface where world models share a latent space with the biological brain.
Recommended reading order. On the first pass, read only Motor Intent Decoding and Landmark System Case Studies — the former makes the kinematic vs. goal-level distinction precise (Shanechi 2024 DPAD is one of the most important recent works), and the latter grounds the abstract pipeline in three real systems: Pitt arm, BrainGate coffee, Walk Again. On a second pass, work through Shared Autonomy (POMDP and hindsight optimization), Hierarchical Planning: BCI + LLM + Robot (BrainBody-LLM / HiCRISP / ROS2 pipeline), and Closed-loop Control and CLDA (co-adaptive online calibration).
Contents of this chapter:
- Motor Intent Decoding — kinematic vs goal-level intent; DPAD (Shanechi 2024)
- Shared Autonomy — the Javdani-Srinivasa POMDP; hindsight optimization; human-machine collaboration in BCI settings
- Hierarchical Planning: BCI + LLM + Robot — BrainBody-LLM, HiCRISP, the ROS2 pipeline
- Closed-loop Control and CLDA — CLDA, SmoothBatch, co-adaptation
- Landmark System Case Studies — Pitt arm (Collinger 2013), BrainGate coffee (Hochberg 2012), Walk Again (Nicolelis 2016)