Efficient Foundation Models
for Real-Time Embodied AI

Making robot foundation models fast enough, cheap enough, and safe enough to actually deploy.

CoRL 2026 · November 12, 2026 · 2:00–6:00 PM CST · Hilton Austin, Room 408

Overview

Embodied foundation models are now good enough to matter. Most of them still miss the latency, memory, and power budgets required to run on the robot itself. This workshop is about that gap: what it actually takes to make an embodied model general, fast, and deployable on real hardware rather than in a datacenter.

This is for you if you work on compression, quantization, distillation, adaptive compute, action chunking, hardware-aware design, edge-cloud systems, or safety and control under timing constraints, whether you build vision-language-action policies, world-action models, or something that does not fit either label. Contributed papers are non-archival; submissions closed October 19 and acceptance notifications have been released.

Why now

RT-2's 55B parameters run at 1 to 3 Hz on cloud TPUs that cost more than the robot. OpenVLA reports roughly 6 Hz on an RTX 4090, and that is a desktop GPU rather than anything that rides onboard. Dexterous manipulation generally wants closed-loop rates well above what these policies deliver, and while some systems close that gap with a fast low-level servo loop under a slower policy, the policy rate still bounds how quickly the robot can change its mind. Meanwhile humanoid programs at scale (Tesla Optimus, Agility Digit, Unitree G1) make per-unit inference cost a first-order constraint: on platforms where the robot costs less than a datacenter GPU, onboard compute is a budget rather than a preference.

Model	Params	Hardware	Reported figure	Measures
Action policies
RT-2	55B	Cloud TPU	1–3 Hz	closed-loop control
OpenVLA	7B	RTX 4090	~6 Hz	throughput
π₀	3.3B	–	50 Hz	open-loop chunk rate
TinyVLA	0.7–1.3B	RTX A6000	20× vs OpenVLA	relative latency
OpenVLA-OFT	7B	A100	109.7 Hz (LIBERO)	throughput
MiniVLA	1B	L40S	12.5 Hz	throughput
HyperVLA	90× fewer active	L4	482→4 ms/step	relative latency
BitVLA	1.58-bit ternary	–	11× vs OFT	relative memory
World models
DreamDojo	2–14B	RTX 5090	10.8 FPS	offline rollout
DreamZero	14B	GB200	7 Hz	closed-loop control

The last column is the point: these numbers measure different things on different hardware, so they cannot be ranked against each other. Building an evaluation that can is one of the things we want papers about.

General

One policy across tasks, objects, and embodiments, rather than a model per skill.

Fast

Closed-loop rates that keep up with contact-rich manipulation.

Deployable

Inside a fixed onboard power, memory, and cost budget, not merely running somewhere.

Systems trade these off differently, and few manage all three at once. Speed and deployability pull apart in particular, because raw Hz can be bought with bigger hardware: OpenVLA-OFT reaches 109.7 Hz, but on an A100 that will never ride on the robot.

What we want papers on

Architecture and compression

Smaller backbones, action chunking, quantization, pruning, distillation, and adaptive compute for policies and world models alike. Results reporting real latency, memory, energy, or control-rate gains on robot-relevant hardware are especially welcome, as is any evidence about whether these techniques compose.

Edge-cloud co-design

Deployment trade-offs across onboard, hybrid, and cloud setups, and across heterogeneous accelerators: GPUs, TPUs, Trainium and Inferentia, mobile NPUs. Total cost of ownership at fleet scale is close to unstudied.

Safety under latency

Interruption, fallback, and certification when inference is slow or its timing is variable. Action chunking commits a robot to open-loop execution with no principled way to interrupt it, and certifying systems with stochastic response time remains open.

Updates

November 6, 2026: Updated workshop session time and venue per the CoRL 2026 workshop schedule: Thursday, November 12, 2026, Afternoon Session (2:00–6:00 PM CST) at Hilton Austin (500 E. 4th St), Room 408.
November 5, 2026: Paper submission results and final acceptance notifications have been released on OpenReview. The final workshop program will be published on November 7, 2026.
October 30, 2026: Launched the EmbodiedX Season 1 Benchmark and Challenge (v1.0.0) on NVIDIA L40S and Google Cloud TPU v6e-1. Checkpoint 1 submissions for the CoRL 2026 live workshop spotlight close November 10, 2026 (23:59 UTC).

Speakers

Jiajun Wu

Stanford University

World models for embodied AI: efficient learned models for online planning.

Yecheng “Jason” Ma

Dyna Robotics

Embodied foundation models on a commercial duty cycle, from DYNA-1 to the DYNA-2 world-action model.

Dhruv Shah

Princeton University and Google DeepMind

Post-training as the stage where most deployable-efficiency decisions are made.

Ouais Alsharif

Google DeepMind

Engineering robot foundation models for production deployment.

Kaifeng Zhang

Sharpa

Topic to be announced.

Additional invited speakers will be announced here as invitations are confirmed.

Schedule

Thursday, November 12, 2026, Afternoon Session (2:00–6:00 PM CST, 14:00–18:00) at Hilton Austin (500 E. 4th Street, Austin, TX 78701), Room 408. The four-hour program pairs four invited talks with contributed lightning talks, a poster session, a moderated panel, and a closing synthesis. Time is deliberately weighted toward the interactive parts, including live profiling on real Jetson hardware.

Time (CST)	Session
14:00	Opening, framing, and live profiling Organizers frame the generality, speed, and deployability trilemma, then profile vanilla OpenVLA on a Jetson on stage, narrating memory and kernel bottlenecks
14:15	Invited talk 30 minutes including questions
14:45	Invited talk 30 minutes including questions
15:15	Lightning talks Five contributed papers, five minutes each, selected for method diversity
15:40	Afternoon break and posters All accepted papers; live demos welcome
16:00	Invited talk 30 minutes including questions
16:30	Invited talk 30 minutes including questions
17:00	Panel, topic to be announced Moderated, seating invited speakers and guests, with an audience vote before and after
17:35	Challenge results Top entries present their submissions
17:45	Synthesis: roadmap working session Organizer-facilitated, feeding the post-workshop report
17:55	Best paper and closing

Room assignments follow the official CoRL 2026 workshop schedule and remain subject to final venue confirmation. Attendees arriving on Thursday should note that conference badge pickup is held exclusively at JW Marriott Austin (110 E. 2nd St, Room 201) from 8:00 AM to 1:30 PM CST and closes before the afternoon session begins.

Call for Papers

We solicit four-page extended abstracts, plus unlimited pages for references, on efficiency across the whole space of embodied foundation models: VLA policies, world-action models, and emerging architectures alike.

Topics

Efficient architectures for VLA policies, world-action models, and emerging designs: action chunking, mixture-of-experts, early exit, diffusion and flow step-reduction
Compression for robot foundation models: quantization, pruning, distillation
Memory-efficient adaptation and fine-tuning
Benchmarking methodology, including whether efficiency techniques compose
Hardware-aware design across accelerators: GPU, TPU, AWS Trainium and Inferentia, mobile NPUs
Edge-cloud hybrid architectures for robot fleets
Safety and latency trade-offs, and fallback policies under stochastic response time
Efficient world models for online planning and real-time safety monitoring

Submission

Four pages plus unlimited references, hosted on OpenReview, reviewed double-blind. Submissions must be anonymized.

Non-archival. Accepted papers are not archived as formal publications, so submitting here does not affect your ability to submit the same work to a main conference. Concurrent submissions and work under review elsewhere are permitted. Previously published work that fits the scope is also eligible, tracked and reviewed separately from new work, so the two do not compete against each other. Every accepted paper in either track gets a poster, and lightning talks are drawn from both. Best Paper is limited to the new-work track. Papers already accepted to the CoRL 2026 main conference are not eligible.

Submission portal: OpenReview. Paper submissions closed on October 19, 2026, and acceptance notifications have been released.

Review and presentation

We aimed for three reviews per submission. With 56 submissions and a program committee of six, each submission received one review, from the organizing committee or the invited program committee, together with a decision rationale from the chair. Paper acceptance decisions have been released on OpenReview as of November 5, 2026. All accepted papers are presented as posters, and five are additionally selected for a five-minute lightning talk.

Embodied Efficiency Challenge

EmbodiedX (embodied-efficiency-bench) is an open, closed-loop serving benchmark charting single-chip latency-throughput Pareto frontiers across Track 1A (π_0.5-LIBERO), Track 1B (GR00T N1.7-3B), and Track 2 (Cosmos-Predict2.5-2B) in both Closed (frozen-checkpoint parity) and Open (Stage-2 task-verified compression) divisions. Season 1 standardizes on two accessible cloud accelerators, NVIDIA L40S (48 GB) and Google Cloud TPU v6e-1 (32 GB), opening unseeded on October 30 with a two-phase commit-reveal protocol that keeps participant kernels private until the freeze.

Awards and checkpoints. Early submissions merged before the November 10, 2026 (23:59 UTC) freeze are recognized live on stage at CoRL on November 12 as Founding Frontier Leaders. Season 1 closes December 18, 2026 (23:59 AoE) for official category awards across each track in Interactive Reactivity (C = 1 p99), Real-Time Fleet Throughput, and Green Embodied Efficiency (Actions/J, Predictions/J), verified on clean organizer hardware over a held-out trace split. Verified Season 1 category winners receive monetary prizes, an official CoRL 2026 EEAI Workshop Award Certificate, an invitation to apply for Google Cloud credits, and an invitation to co-author the post-season EmbodiedX Season 1 Systems Report.

Challenge page and live leaderboard → · Awards & dates · Protocol & commit-reveal · GitHub repository · Methodology specification (PDF)

Important Dates

October 19, 2026: Paper submission deadline (extended from October 12)
October 30, 2026: EmbodiedX Season 1 benchmark and leaderboard open
November 5, 2026: Final paper acceptance notifications released on OpenReview
November 7, 2026: Final workshop program published
November 10, 2026: Challenge Checkpoint 1: CoRL live spotlight freeze (23:59 UTC)
November 12, 2026: Workshop at CoRL 2026 (2:00–6:00 PM CST, Hilton Austin, Room 408), with live Founding Frontier presentation
December 18, 2026: Challenge Checkpoint 2: Season 1 awards freeze (23:59 AoE; +24h private finalist container reveal)

Paper deadlines are 23:59 Anywhere on Earth (AoE). Following the December 18 Season 1 freeze, the public EmbodiedX leaderboard remains open continuously for rolling pull requests in 2027.

Organizers

Jiafan Yu

Google

Inference optimization for embodied foundation models.

Ouais Alsharif

Google DeepMind

Engineering robot foundation models for production deployment.

Changliu Liu

Carnegie Mellon University

Safe robot learning and real-time optimization-based control.

Dhruv Shah

Princeton University and Google DeepMind

Post-training for robot foundation models, and navigation policies on constrained hardware.

Jiajun Wu

Stanford University

World models and visual reasoning for robot learning.

Zhuoyang Zhang

MIT

Efficient vision-language-action models and foundation-model acceleration.

Program Committee

We thank the following reviewers for volunteering their time on the contributed papers.

Aaryan Agrawal, Service Robot Co.
Yoofi Annan, Figure AI
Rajasekhar Josyula, Tesla AI
Nitheesh K Lakshminarayana, Nimble Robotics
Mangat Rai, IBM
Sriharsha Sammeta, Google