Language as Robot Middleware
Andy Zeng
Princeton Robotics Seminar
Manipulation
amazing skill! teppanyaki steak master
https://youtu.be/5qVMIKCn_Cs
Problem difficulty = #DoFs (robot) + #DoFs (environment)
Robotics as a platform to study intelligent machines
Robotics as a platform to study intelligent machines
SHRDLU, Terry Winograd, MIT 1968
"will you please stack up both of the red blocks and either a green cube or a pyramid"
person:
"what does the box contain?"
person:
"the blue pyramid and blue block."
robot:
"can the table pick up blocks?"
person:
"no."
robot:
"why did you do that?"
person:
"ok."
robot:
"because you asked me to."
robot:
Simple things are surprisingly complex
"At an analytical level, pushing is a well understood problem... These are usually based on Coulomb's friction law..."
"The reality, however, is bitter... the sensitivity of the task to small changes in contact geometry, along with the variability of friction, hinders accurate predictions."
More than a Million Ways to Be Pushed. A High-Fidelity Experimental Dataset of Planar Pushing
Kuan-Ting Yu, Maria Bauza, Nima Fazeli, Alberto Rodriguez, IROS 2016
Imitation Learning with Behavior Cloning
Supervised Learning
1. Collect dataset with human teleop
2. Train end-to-end deep networks
- High-rate (velocity) control
- Explicit (e.g. MSE) losses
Transformers
or ConvNets
Actions
Pixels
Imitation Learning with Behavior Cloning
Supervised Learning
1. Collect dataset with human teleop
2. Train end-to-end deep networks
- Data hungry
- Policy often gets "stuck"
- High-rate (velocity) control
- Explicit (e.g. MSE) losses
In Practice
Transformers
or ConvNets
Actions
Pixels
Imitation Learning with Behavior Cloning
Supervised Learning
1. Collect dataset with human teleop
2. Train end-to-end deep networks
- Data hungry
- Policy often gets "stuck"
- High-rate (velocity) control
- Explicit (e.g. MSE) losses
In Practice
Transformers
or ConvNets
Actions
Pixels
Imitation Learning with Behavior Cloning
- Data hungry
- Policy often gets "stuck"
In Practice
Implicit Behavior Cloning
Implicit Behavioral Cloning, CoRL 2021
Pete Florence, Corey Lynch, Andy Zeng, Oscar Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, Jonathan Tompson
Implicit Behavior Cloning
LInfoNCE=∑i=1N−log(p~θ(yi∣ x, {y~ij}j=1Nneg.))
\mathcal{L}_{\text{InfoNCE}} = \sum_{i=1}^N -\log \big( \tilde{p}_{\theta}( {\color{black} \mathbf{y}_i} | \ \mathbf{x}, \ {\color{red}\{\tilde{\mathbf{y}}^j_i\}_{j=1}^{N_{\text{neg.}}} } ) \big)
p~θ(yi∣ x, {y~ij}j=1Nneg.)=e−Eθ(xi,yi)+∑j=1Nnege−Eθ(xi,y~ij)e−Eθ(xi,yi)
\tilde{p}_{\theta}( {\color{black} \mathbf{y}_i} | \ \mathbf{x}, \ {\color{red}\{\tilde{\mathbf{y}}^j_i\}_{j=1}^{N_{\text{neg.}}} } ) = \frac{e^{-E_{\theta}(\mathbf{x}_i, {\color{black} \mathbf{y}_i} )}} {e^{-E_{\theta}( \mathbf{x}_i, {\color{black} \mathbf{y}_i})} + {\color{red} \sum_{j=1}^{N_{\text{neg}}}} e^{-E_{\theta}(\mathbf{x}_i, {\color{red} \tilde{\mathbf{y}}^j_i} )} }
Implicit Behavioral Cloning, CoRL 2021
Pete Florence, Corey Lynch, Andy Zeng, Oscar Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, Jonathan Tompson
Implicit Behavior Cloning
LInfoNCE=∑i=1N−log(p~θ(yi∣ x, {y~ij}j=1Nneg.))
\mathcal{L}_{\text{InfoNCE}} = \sum_{i=1}^N -\log \big( \tilde{p}_{\theta}( {\color{black} \mathbf{y}_i} | \ \mathbf{x}, \ {\color{red}\{\tilde{\mathbf{y}}^j_i\}_{j=1}^{N_{\text{neg.}}} } ) \big)
p~θ(yi∣ x, {y~ij}j=1Nneg.)=e−Eθ(xi,yi)+∑j=1Nnege−Eθ(xi,y~ij)e−Eθ(xi,yi)
\tilde{p}_{\theta}( {\color{black} \mathbf{y}_i} | \ \mathbf{x}, \ {\color{red}\{\tilde{\mathbf{y}}^j_i\}_{j=1}^{N_{\text{neg.}}} } ) = \frac{e^{-E_{\theta}(\mathbf{x}_i, {\color{black} \mathbf{y}_i} )}} {e^{-E_{\theta}( \mathbf{x}_i, {\color{black} \mathbf{y}_i})} + {\color{red} \sum_{j=1}^{N_{\text{neg}}}} e^{-E_{\theta}(\mathbf{x}_i, {\color{red} \tilde{\mathbf{y}}^j_i} )} }
Implicit Behavioral Cloning, CoRL 2021
Pete Florence, Corey Lynch, Andy Zeng, Oscar Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, Jonathan Tompson
Implicit Behavior Cloning & Transporter Nets
LInfoNCE=∑i=1N−log(p~θ(yi∣ x, {y~ij}j=1Nneg.))
\mathcal{L}_{\text{InfoNCE}} = \sum_{i=1}^N -\log \big( \tilde{p}_{\theta}( {\color{black} \mathbf{y}_i} | \ \mathbf{x}, \ {\color{red}\{\tilde{\mathbf{y}}^j_i\}_{j=1}^{N_{\text{neg.}}} } ) \big)
p~θ(yi∣ x, {y~ij}j=1Nneg.)=e−Eθ(xi,yi)+∑j=1Nnege−Eθ(xi,y~ij)e−Eθ(xi,yi)
\tilde{p}_{\theta}( {\color{black} \mathbf{y}_i} | \ \mathbf{x}, \ {\color{red}\{\tilde{\mathbf{y}}^j_i\}_{j=1}^{N_{\text{neg.}}} } ) = \frac{e^{-E_{\theta}(\mathbf{x}_i, {\color{black} \mathbf{y}_i} )}} {e^{-E_{\theta}( \mathbf{x}_i, {\color{black} \mathbf{y}_i})} + {\color{red} \sum_{j=1}^{N_{\text{neg}}}} e^{-E_{\theta}(\mathbf{x}_i, {\color{red} \tilde{\mathbf{y}}^j_i} )} }
Transporter Networks: Rearranging the Visual World for Robotic Manipulation, CoRL 2020
Andy Zeng, Pete Florence, Jonathan Tompson, Stefan Welker, Jonathan Chien, Maria Attarian, Travis Armstrong, Ivan Krasin, Dan Duong, Ayzaan Wahid, Vikas Sindhwani, Johnny Lee
A generalization of loss functions from Transporter Nets (spatial action maps)
Implicit Behavior Cloning
Implicit Behavioral Cloning, CoRL 2021
Pete Florence, Corey Lynch, Andy Zeng, Oscar Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, Jonathan Tompson
Implicit Behavior Cloning
Implicit Behavioral Cloning, CoRL 2021
Pete Florence, Corey Lynch, Andy Zeng, Oscar Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, Jonathan Tompson
Implicit Behavior Cloning
Implicit Behavioral Cloning, CoRL 2021
Pete Florence, Corey Lynch, Andy Zeng, Oscar Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, Jonathan Tompson
Interpolation vs Extrapolation*
- Sample efficiency improvements
Deep Learning is a Box
Interpolation
Extrapolation
Deep Learning is a Box
Interpolation
Extrapolation
Roboticist
Vision
NLP
Deep Learning is a Box
Interpolation
Extrapolation
Internet
Meanwhile in NLP...
Large Language Models
Large Language Models
Internet
Meanwhile in NLP...
Books
Recipes
Code
News
Articles
Dialogue
Demo
Quick Primer on Language Models
Tokens (inputs & outputs)
Transformers (models)
Attention Is All You Need, NeurIPS 2017
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin
Quick Primer on Language Models
Tokens (inputs & outputs)
Transformers (models)
Pieces of words (BPE encoding)
big
bigger
per word:
biggest
small
smaller
smallest
big
er
per token:
est
small
Attention Is All You Need, NeurIPS 2017
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin
Quick Primer on Language Models
Tokens (inputs & outputs)
Transformers (models)
Self-Attention
Pieces of words (BPE encoding)
big
bigger
per word:
biggest
small
smaller
smallest
big
er
per token:
est
small
y
y
x1
x_1
x2
x_2
x3
x_3
Attention Is All You Need, NeurIPS 2017
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin
softmax(dkQK⊺)V
\text{softmax}(\frac{QK^\intercal}{\sqrt{d_k}})V
Bigger is Better
Neural Language Models: Bigger is Better, WeCNLP 2018
Noam Shazeer
Bigger is Better
Neural Language Models: Bigger is Better, WeCNLP 2018
Noam Shazeer
Bigger is Better
Neural Language Models: Bigger is Better, WeCNLP 2018
Noam Shazeer
Bigger is Better
Neural Language Models: Bigger is Better, WeCNLP 2018
Noam Shazeer
Bigger is Better
Neural Language Models: Bigger is Better, WeCNLP 2018
Noam Shazeer
Somewhere in the space of interpolation
Example?
Lives robot planning
Large Language Models on Robots
Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language
https://socraticmodels.github.io
Open research problem! but here's one way to do it...
Do As I Can, Not As I Say: Grounding Language in Robotic Affordances
https://say-can.github.io
Large Language Models on Robots
Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language
https://socraticmodels.github.io
Open research problem! but here's one way to do it...
Visual Language Model
CLIP, ALIGN, LiT,
SimVLM, ViLD, MDETR
Human input (task)
Do As I Can, Not As I Say: Grounding Language in Robotic Affordances
https://say-can.github.io
Large Language Models on Robots
Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language
https://socraticmodels.github.io
Open research problem! but here's one way to do it...
Visual Language Model
CLIP, ALIGN, LiT,
SimVLM, ViLD, MDETR
Human input (task)
Large Language Model for Planning (e.g. SayCan)
Language-conditioned Policies
Do As I Can, Not As I Say: Grounding Language in Robotic Affordances
https://say-can.github.io
Live Demo
Language as a "state" representation
3D Recontruction
Language as a "state" representation
z
z
semantic?
compact?
compositional?
general?
interpretable?
Perception
Actions
Planning
3D Recontruction
Language as a "state" representation
z
z
semantic?
compact?
compositional?
general?
interpretable?
Perception
Actions
Planning
3D Recontruction
scope of guarantees
safety & bias mitigation
Language as a "state" representation
3D Recontruction
z
z
Perception
Actions
Planning
semantic? ✓
compact? ✓
compositional? ✓
general? ✓
interpretable? ✓
What about language?
Ego4D
Limits of language as a "state" representation
- Loses spatial precision
- Highly multimodal (lots of different ways to say the same thing)
- Not as information-rich as in-domain representations (e.g. images)
Limits of language as a "state" representation
- Only for high level? what about control?
Perception
Planning
Control
Socratic Models
Inner Monologue
ALM + LLM + VLM
SayCan
Wenlong Huang et al, 2022
LLM
Imitation? RL?
Engineered?
Intuition and commonsense is not just a high-level thing
Intuition and commonsense is not just a high-level thing
Applies to low-level behaviors too
Is the "dark matter" of robotics
- Spatial: "move a little bit to the left"
- Temporal: "move faster"
- Functional: "balance yourself"
Demo
Intuition and commonsense is not just a high-level thing
Seems to be stored in the depths of in language models... how to extract it?
Applies to low-level behaviors too
Is the "dark matter" of robotics
- Spatial: "move a little bit to the left"
- Temporal: "move faster"
- Functional: "balance yourself"
Language models can write code
Code as a medium to express low-level commonsense
Live Demo
Language models can write code
Code as a medium to express low-level commonsense
Live Demo
Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, Andy Zeng
code-as-policies.github.io
Code as Policies: Language Model Programs for Embodied Control
Language models can write code
Code as a medium to express low-level commonsense
Live Demo
Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, Andy Zeng
code-as-policies.github.io
Code as Policies: Language Model Programs for Embodied Control
Language models can write code
Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, Andy Zeng
code-as-policies.github.io
Code as Policies: Language Model Programs for Embodied Control
use NumPy,
SciPy code...
Language models can write code
Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, Andy Zeng
code-as-policies.github.io
Code as Policies: Language Model Programs for Embodied Control
- PD controllers
- impedance controllers
Language models can write code
Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, Andy Zeng
code-as-policies.github.io
Code as Policies: Language Model Programs for Embodied Control
Language models can write code
Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, Andy Zeng
code-as-policies.github.io
Code as Policies: Language Model Programs for Embodied Control
What is the foundation models for robotics?
How much data do we need?
Robot Learning
Language Models
Not a lot of robot data
Lots of Internet data
How much data do we need?
Robot Learning
Language Models
Not a lot of robot data
Lots of Internet data
How much data do we need?
Robot Learning
Language Models
Not a lot of robot data
Lots of Internet data
50 expert demos
How much data do we need?
Robot Learning
Language Models
Not a lot of robot data
Lots of Internet data
500 expert demos
50 expert demos
How much data do we need?
Robot Learning
Language Models
Not a lot of robot data
Lots of Internet data
500 expert demos
5000 expert demos
50 expert demos
How much data do we need?
Robot Learning
Language Models
Not a lot of robot data
Lots of Internet data
500 expert demos
5000 expert demos
50 expert demos
How much data do we need?
Robot Learning
Language Models
- Finding other sources of data (sim, YouTube)
- Improve data efficiency with prior knowledge
Not a lot of robot data
Lots of Internet data
500 expert demos
5000 expert demos
50 expert demos
Scale alone might not be enough
Robot Learning
Language Models
Not a lot of robot data
Lots of Internet data
adapted from Tomás Lozano-Pérez
Machine learning is a box
... but robotics is a line
- Finding other sources of data (sim, YouTube)
- Improve data efficiency with prior knowledge
Different embodiments etc....
A possible middleground
Robot Learning
Language Models
Not a lot of robot data
Lots of Internet data
Embrace Compositionality
adapted from Tomás Lozano-Pérez
Deep learning is a box
... but robotics is a line
A possible middleground
Robot Learning
Language Models
Not a lot of robot data
Lots of Internet data
Embrace Compositionality
adapted from Tomás Lozano-Pérez
Deep learning is a box
... but robotics is a line
2. composing them
autonomously
1. build boxes
Towards grounding everything in language
Language
Control
Vision
Tactile
Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language
https://socraticmodels.github.io
"Language" as the glue for robots & AI
Language
Perception
Planning
Control
Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language
https://socraticmodels.github.io
"Language" as the glue for robots & AI
Language
Perception
Planning
Control
Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language
https://socraticmodels.github.io
We have some reason to believe that
"the structure of language is the structure of generalization"
To understand language is to understand generalization
https://evjang.com/2021/12/17/lang-generalization.html
Sapir–Whorf hypothesis
Towards grounding everything in language
Language
Perception
Planning
Control
Humans
Towards grounding everything in language
Language
Perception
Planning
Control
Humans
Not just for general robots,
but for human-centered intelligent machines!
- Grounding safety and risk in language
Things I wish my robot had...
- Grounding safety and risk in language
- Compositional guarantees on safety and generalization?
Things I wish my robot had...
- Grounding safety and risk in language
- Compositional guarantees on safety and generalization?
- Embed more LLM prior knowledge into control?
Things I wish my robot had...
- Grounding safety and risk in language
- Compositional guarantees on safety and generalization?
- Embed more LLM prior knowledge into control?
- PAC-bayes on LLM planners?
Things I wish my robot had...
Thank you!
Pete Florence
Adrian Wong
Johnny Lee
Vikas Sindhwani
Stefan Welker
Vincent Vanhoucke
Kevin Zakka
Michael Ryoo
Maria Attarian
Brian Ichter
Krzysztof Choromanski
Federico Tombari
Jacky Liang
Aveek Purohit
Wenlong Huang
Fei Xia
Peng Xu
Karol Hausman
and many others!
Amazon Picking Challenge
arc.cs.princeton.edu
Team MIT-Princeton
excel at simple things, adapt to hard things
How to endow robots with
"intuition" and "commonsense"?
Language as Robot Middleware Andy Zeng Princeton Robotics Seminar
2022-Princeton-Talk
By Andy Zeng
2022-Princeton-Talk
- 549
