I really like the MPU-6050. It's a small, 4x4mm QFN chip that has an accelerometer, gyroscope, and a legendary "DMP" that we'll get to in a minute.

But first, why do we need spatial data?

At first, my goal was to give the wearable device a bit of contextual information: the user is probably not using the device 24/7, and asking them to turn it off (which currently amounts to pulling out the battery :) ) isn't feasible. So the device needs to be able to sleep at times when reporting distance is useless (if the device is sitting still, potentially not even worn) or annoying (vibration when the device is pointed to the floor while walking.) 

Now, Milestone 1 (heretofore MS1? Please?) has a "dumb" spatial response: if the device is unmoved for >2 minutes, OR if it is pointed more than 60 degrees below the horizon, it enters sleep mode.

But there are potentially some use cases where these situations are a part of normal operation: pointing at steps, for example, would not be possible with this blanket rule. 

I'd like to detect discrete gesture patterns, and define device behavior that way. For example, detecting a natural walking gait, and sleeping only during the bottom sweep of the swing. 

This would also be rather important in prolonging battery life, as the motor is by far the most significant power consumer. 

In any case, that's where I'm headed, but I'll need to wrap my head around those beautifully complex quaternions returned by Jeff Rowberg's MPU-6050 library first.

Also, sourcing expensive sensors from China often ends poorly :(. My current boards do not work with the raw MPU-6050 chips coming from China; they return wildly unstable values but otherwise expose the right data structures and I2C commands. It remains to be seen whether this is a fault with the ICs, or the layout itself (lots of passives to support the chip!).