GoPiGo Robot More Fun Than Facebook's Robot

[cyclicalobsessive]

GoPiGo Robot More Fun Than Facebook’s Robot

Hypothesis: A Modular Robotics / Dexter Industries GoPiGo3 robot with a single Raspberry Pi 3B central processing unit and the PiCamera, should be sufficient to create a fun robot.

Comparative:
Facebook Robot: “Near-perfect point-goal navigation from 2.5 billion frames of experience”[1]

• “even failing 1 out of 100 times is not acceptable in the physical world”
• error might damage robot or surroundings
• learns to navigate without a map
• “maps become outdated the moment they are created. Buildings change”
• “no scope for mistakes of any kind – no wrong turn …, no exploration”
• agent learns to exploit statistical regularities in real indoor environments
• distributed GPU processes, collectors (simulated actors with sensors)
• 2.5B steps or 80 years experience
• three days wall clock using 64 GPU each having 10K fps processing ability
• 90% success in 1 day with 8 GPU processed 100M steps
• RGB-D imagery with compass and GPS
• Cost? Probably over $100K
• Note: RGB-D without GPS and compass failed at 100M steps, 16% success at 2.5B
• Success is defined as “near-perfect point-goal navigation”
  .
  .

Suggested GoPiGo Fun Robot Principles:

• Mistakes are expected and tolerated
• Building does not change (quickly)
• Everything is an estimate
• There is no deadline
• Use what you have available
• Learn Environment Map from images
• Parameterize map using encoder data
• Increase trust/adjust with ToF Distance Sensor and IMU
• Increase trust with multiple measurements over time
• Learn Calibrations (Image, Encoders, ToF Distance Sensor, IMU)
• Motivation for map is to allow exploration beyond sight of recharge dock
• Map and exploration are actually much less important than human interaction
• Cost? around $250-$350 (GoPiGo3, Raspberry Pi, PiCam, Battery and Charger)
  [ToF Distance Sensor, Servo Kit, IMU]
• Success: “robot is fun”

.
.
GoPiGo Pertinent Parameters:

Raspberry Pi 3B:

• CPU 4x Cortex-A53 64-bit 1.2GHz
• GPU 28.8 GFLOPS
• RAM 1GB shared CPU/GPU memory

.
.
Power System:

• 9.6v nominal rechargeable battery provides 1 to 8 hour “life segments”

.
.

References:
[1] https://ai.facebook.com/blog/near-perfect-point-goal-navigation-from-25-billion-frames-of-experience/

[jimrh]

So, let me see If I have this right:
I would have the choice of either 2 and a half BMW/Mercedes or ONE Facebook robot?

Additional data points:

• You don’t have to buy the full kit all at once since there’s still a lot of very useful information, (make that read “fun”), you can get / have with just the basic kit.
• You can purchase additional features as/when you can afford and/or need them. If you don’t need the ICBM launcher, you can save for something else, like maybe the “oil-spill skimmer” attachment?
• If there’s a feature or attachment that Dexter doesn’t offer, being based on the Pi makes it possible to find what you need from other sources too.
  • No one has it? Make it yourself! Screws, bits of metal or plastic, buttons and switches, glue, all can be used and all are easy to find and inexpensive.
• You don’t have to commit to a particular line of research as the GoPiGo (et. al.) Dexter 'bots are virtually infinitely variable in what they can do.
  • Carl is a good example of research on “autonomous robotics” using a limited amount of AI.
  • There’s another 'bot with a series of experiments towards becoming an autonomous fire-fighting robot.
  • My own 'bot is providing me with a far better education in applied robotics than any university could hope to do, and for a LOT less money!
• Because both the 'bot itself and any accessories are relatively inexpensive, (bumpers made out of 12-cent-each kabab skewers and two buttons anyone?), you don’t have to be as afraid of failure. Even a full-blown Pi-4 with 4 gigs of memory is right at $50 US - and you can do excellently with far less.
• Since it’s made out of (relatively) thick and durable plastic, it can take a lot of punishment and come back for more.
  • This is especially important for me since I have the dual challenges of both piss-poor coordination and a tremor that would make me illegal on the West Coast. “Sorry!” Charlie isn’t just a euphemism; I have a carpeted floor for a reason.
  • The logical extension of that is the use of robots like the GoPiGo in classrooms where the students are more physically challenged without the fear of destroying zillions of dollars worth of equipment. Try THAT with hundred-grand-a-pop Facebook robots!
• As I have said before, despite the fact that the GoPiGo is a comparatively inexpensive robot, that doesn’t mean that it’s just a toy. It has some real chops and the ability to go wherever your imagination is willing to take it.
• If you and/or your students are too intimidated by the GoPiGo, Dexter has the GiggleBot which, though simpler and less intimidating, is still an excellent 'bot in its own right and has a lot of potential for both fun and learning. My own granddaughters gave the GiggleBot a big “Two micro:bits up!” (And I am glad I had the foresight to buy one for each of them!)
  .

There are a lot of very interesting 'bots out there but most are either way too expensive or are merely “toy” class 'bots with limited potential.

Though my experience with robotics is limited, my experience with electronics, electronic design and fabrication for both civilian and military devices isn’t.

Based on this experience, gained over decades, I feel it is safe to say that the GoPiGo has an important and valuable place in robotics education and training.

What say ye?

Jim “JR”

[jimrh]

Success: “robot is fun”

+10,000,000

More people/educators should take this attitude!

[jimrh]

We propose a simple, synchronous, distributed RL method that scales well. We call this method decentralized distributed proximal policy optimization, as it is decentralized (has no parameter server) and distributed (runs across many different machines), and we use it to scale proximal policy optimization, a previously developed technique ( Schulman et al., 2017 ). In DD-PPO, each worker alternates between collecting experience in a resource-intensive, GPU-accelerated simulated environment and then optimizing the model. This distribution is synchronous — there is an explicit communication stage in which workers synchronize their updates to the model.

Another point:

In Facebook’s model, (part of which is quoted above), they obviously envision a large distributed network of systems, (or one huge multi-processor or multi-GPU system), to do the necessary deep learning needed to run that beastie.

Unless you “just happen” to have Facebook’s deep pockets, (or massive government/university funding), no one is going to be able to approach that kind of scaling.

Part of the fun of working with a, (ahem!), “lesser system” is the need to be both clever and efficient since we don’t have either, (comparatively), infinite processing power or funding.

Though university-level processing research has its place, 'bots like the GoPiGo provide valuable information and experience in a more “real-world” environment.

If we assume that you are trying to develop a marketable robotics-based solution to some problem, it’s a virtual lead-pipe-cinch that you won’t have that kind of budget. Likewise, your customers aren’t going to pay mega-bucks for a household appliance either.

This is also applicable to lower-budget research and/or educational projects. 'Bots like the GoPiGo allow those with a more limited budget to explore questions and propose solutions using a platform that is powerful enough to do effective research without requiring unreasonable amounts of money.

Because of this, systems like the GoPiGo are invaluable research tools; not to mention being a lot of fun!