Geng Chen (@gengchen) profile picture

Ph.D. student at UPenn @GRASPlab | Prev Master student @UCSDJacobs, Undergrad @sjtu1896 | Working on Embodied AI & Robot Learning.

Philadelphia, PA
Joined May 2022
Photo by Geng Chen (@gengchen)Photo by Geng Chen (@gengchen)Photo by Geng Chen (@gengchen)Photo by Geng Chen (@gengchen)Photo by Geng Chen (@gengchen)Photo by Geng Chen (@gengchen)Photo by Geng Chen (@gengchen)Photo by Geng Chen (@gengchen)Photo by Geng Chen (@gengchen)Photo by Geng Chen (@gengchen)
Geng Chen retweeted
RoboSkills agent-as-policy is now in the real world! ๐Ÿค– - Hard coordination: twisting the lid off a bottle. - Precise action: picking a slim pen and stacking, with no crashes. - Cross embodiment: one workflow for both grippers and hands! code: github.com/openrobokit/robosโ€ฆ
2
7
1
22
1,505
Geng Chen retweeted
Our G1 robot found its way to the pantry, picked up a Pringles can, and handed it over. No pre-built map. No training for this task. Behind it is a coding agent, writing, running, and fixing the robot's programs as it goes. And it self-evolves: what works is saved as code for the next task, no retraining needed. This is ASENA. Joint work with @Isabella__Liu, @edmund_bu, Johan Bjorck, @yin_hongxu, @zhengyiluo, @jankautz, @DrJimFan, @yukez, and @Sifei30488L. ๐Ÿงต๐Ÿ‘‡
11
30
8
139
16,820
Can we scale robot learning without scaling human labor? ๐Ÿ“ˆ Demonstrations, resets, failure recovery, constant supervisionโ€”collecting robot data still takes substantial human effort. ๐Ÿ”‘ Three keys let agents take on more of the work: 1. Real2Sim2Real: Agents explore and learn in digital twins, where trial and error is safer and cheaper. 2. Autonomous Recovery: Robots learn to recover from unexpected failures instead of waiting for human resets. 3. Human-on-Demand Learning: Multiple robots run in parallel, and humans step in only when autonomous recovery falls short. ๐Ÿฅ‘ Recova brings these together to get more robot learning out of every hour of human effort. research.nvidia.com/labs/geaโ€ฆ Joint work with @anjjei, Johan Bjorck, @ZhidingYu, @yin_hongxu, @jankautz, @DrJimFan, @yukez, and @Sifei30488L. ๐Ÿ‘‡ Here's how it works ๐Ÿงต
5
31
11
159
17,962
When we started HARMONY six months ago, we were curious how far agentic approaches could take 3D reconstruction. The field has advanced quickly, and we see strong potential in combining high-level reasoning with geometric optimization. Weโ€™re releasing an open-source framework competitive with Astra in reconstruction quality, together with the HARMONY300 benchmark. Proud of @sofine_sun and @chenwangcw!
Excited to share our new ๐—ผ๐—ฝ๐—ฒ๐—ป-๐˜€๐—ผ๐˜‚๐—ฟ๐—ฐ๐—ฒ work, ๐—›๐—”๐—ฅ๐— ๐—ข๐—ก๐—ฌ, a framework for reconstructing compositional 3D scenes from single indoor images. Much like a 3D artist, our VLM agent builds the scene in a ๐—ต๐—ถ๐—ฒ๐—ฟ๐—ฎ๐—ฟ๐—ฐ๐—ต๐—ถ๐—ฐ๐—ฎ๐—น ๐—ผ๐—ฟ๐—ฑ๐—ฒ๐—ฟ: first the room structure and camera, then wall-mounted objects, free-standing furniture, and finally dependent decorations. After each stage, it compares its render against the input image, reflects on what went wrong, and iteratively corrects the scene to prevent errors from accumulating. Even with older VLM backbones, our framework remains ๐—ฐ๐—ผ๐—บ๐—ฝ๐—ฒ๐˜๐—ถ๐˜๐—ถ๐˜ƒ๐—ฒ ๐˜„๐—ถ๐˜๐—ต ๐—š๐—ฃ๐—ง-๐Ÿฒ ๐—”๐˜€๐˜๐—ฟ๐—ฎ ๐—ถ๐—ป ๐˜ƒ๐—ถ๐˜€๐˜‚๐—ฎ๐—น ๐—พ๐˜‚๐—ฎ๐—น๐—ถ๐˜๐˜†, ๐˜„๐—ต๐—ถ๐—น๐—ฒ ๐—ฎ๐—ฐ๐—ต๐—ถ๐—ฒ๐˜ƒ๐—ถ๐—ป๐—ด ๐—ฏ๐—ฒ๐˜๐˜๐—ฒ๐—ฟ ๐—ด๐—ฒ๐—ผ๐—บ๐—ฒ๐˜๐—ฟ๐—ถ๐—ฐ ๐—ฎ๐—น๐—ถ๐—ด๐—ป๐—บ๐—ฒ๐—ป๐˜ for single-image 3D scene reconstruction. We also release ๐—›๐—”๐—ฅ๐— ๐—ข๐—ก๐—ฌ๐Ÿฏ๐Ÿฌ๐Ÿฌ, a benchmark for single-image 3D scene reconstruction with 300 scenes across three difficulty tiers: easy, medium, and complicated. ๐Ÿ”— Project: cwchenwang.github.io/harmony ๐Ÿ“Š Leaderboard: cwchenwang.github.io/harmonyโ€ฆ ๐Ÿค— Data: huggingface.co/datasets/Shufโ€ฆ ๐Ÿ“„ Paper: arxiv.org/abs/2609.26793
2
7
40
6,730
Children learn to walk by falling. Robots fail tooโ€”but we usually throw those failures away. Yet failed rollouts tell us something successful demonstrations cannot: what bad actions actually lead to. So, can robots learn from their failures too? Introducing FACT: Failure-Aware Causal Training for World-Action Models. ๐Ÿ”—website: fact-wam.github.io/ ๐Ÿ“„arxiv: arxiv.org/abs/2608.10232 ๐Ÿ’ป code: github.com/Bariona/FACT
15
43
10
323
77,627
Geng Chen retweeted
Human demos contain rich hand-object skills. The challenge is the embodiment gap. Human and robot hands differ in shape, joints, and contacts. Introducing ConTrack, our latest work that turns contact-rich human hand-object demos into robot motions.
6
31
3
170
26,621
Geng Chen retweeted
#CVPR2026 GR3D ๐ŸงŠ โ€” A single VLM that grounds in 2D, grounds in 3D, and reasons with visual chain-of-thought โ€” all at once. Excited to share our paper, Grounded 3D-Aware Spatial Vision-Language Modeling!
4
36
3
326
35,628
Ever want to have a single policy to control diverse robots as well as different dexterous hands, or to observe the emergent behavior under cross embodiment training? Introducing our #CVPR2026 paper XL-VLA, Cross-Hand Latent Representation for Vision-Language-Action Models.
11
42
20
277
91,203
Geng Chen retweeted
Learning from vast video data allows point-flow-based planners to create generalizable task plans, guiding robots via future keypoint trajectories. But how can we ensure low-level execution doesn't become the system's bottleneck? Introducing our #ICLR2026 paper, HinFlow: Translating Flow to Policy via Hindsight Online Imitation
3
15
7
94
26,578
"Cross-embodiment" is a sign of generalization. Weโ€™ve seen huge progress in manipulation and navigation โ€” but what about humanoid whole-body control? Can ONE policy control multiple different humanoids? Meet our #ICRA2026 work ๐Ÿฆ…EAGLE: Embodiment-Aware Generalist Specialist Distillation for Unified Humanoid Whole-Body Control. Instead of brute-force URDF / morphology domain randomization, we iteratively distill specialists into one generalist. We also find that embodiment-aware representations matter for policy learning. ๐Ÿ”— website: eagle-wbc.github.io/ ๐Ÿ“œ arXiv: arxiv.org/abs/2602.02960
6
34
7
175
30,706
VLA models are boomingโ€”but one fundamental question is yet seldom answered: How does the choice of the base VLM affect VLA performanceโ‰๏ธ We run a large-scale systematic study to answer this, in collaboration with @Alibaba_Qwen cladernyjorn.github.io/VLM4Vโ€ฆ
13
57
8
443
41,338
How far can we push the limit of in-hand manipulation dexterity? Introducing our work on motion capture: DexterCap & DexterHand ! DexterCap: A high-fidelity motion capture system for intricate in-hand manipulation motion. DexterHand: A dataset featuring true in-hand dexterity, reorientation, finger gaiting, and even manipulating a Rubik's Cube like a speedcuber ! ๐Ÿงฉ - Project Page: pku-mocca.github.io/Dextercaโ€ฆ - Experience it via our online interactive visualization: lyt0112.com/projects/DexterCโ€ฆ #Animation #CharacterAnimation #MotionCapture #Graphics #EmbodiedAI #DexterousManipulation
2
9
4
38
20,121
What makes robot hands dexterous? ๐Ÿค–๐Ÿ–๏ธ We generate robot hands + control, sim-to-real, in under 24 hours. Paper: Cross-Embodied Co-Design for Dexterous Hands ๐Ÿ”ฅ Rapid evaluation w/ cross-embodied policies ๐Ÿฆพ Open-source modular hand platform ๐Ÿ’ก Automated full-hand generation
7
32
5
153
21,286
Geng Chen retweeted
Meet ACE-F โ€” a novel, foldable teleoperation platform for collecting high-quality robot demonstration data across robot embodiments. Using a specialized soft-controller pipeline, we interpret end-effector positional deviations as virtual force signals to provide the user with force feedback, without requiring expensive sensors! ACE-F simplifies control for a diverse array of robot platforms, making complex tasks that require dexterous manipulation highly intuitive. Check out the project website here! acefoldable.github.io/
2
14
2
27
5,710
How do you teach a robot to do something it has never seen before? ๐Ÿค– With human data. Our new Human0 model is co-trained on human and humanoid data. It allows the robot to understand a novel language command and execute it perfectly in the wild without prior practice. Real-world success rate: ๏ฝž100%. Watch it happen ๐Ÿ‘‡
A large human behavior model. Introducing In-N-On, our latest findings in scaling egocentric data for humanoids. 1. Pre-training and post-training with human data 2. 1,000+ hours of in-the-wild data and 20+ hours of on-task data with accurate action labels Website: xiongyicai.github.io/In-N-Onโ€ฆ Arxiv: arxiv.org/abs/2511.15704 By simply scaling data, our robot can follow novel language instruction. Check out the ๐Ÿงต
5
27
4
174
22,353
A large human behavior model. Introducing In-N-On, our latest findings in scaling egocentric data for humanoids. 1. Pre-training and post-training with human data 2. 1,000+ hours of in-the-wild data and 20+ hours of on-task data with accurate action labels Website: xiongyicai.github.io/In-N-Onโ€ฆ Arxiv: arxiv.org/abs/2511.15704 By simply scaling data, our robot can follow novel language instruction. Check out the ๐Ÿงต
5
47
11
232
80,580
Geng Chen retweeted
Meet BFM-Zero: A Promptable Humanoid Behavioral Foundation Model w/ Unsupervised RL๐Ÿ‘‰ lecar-lab.github.io/BFM-Zeroโ€ฆ ๐ŸงฉONE latent space for ALL tasks โšกZero-shot goal reaching, tracking, and reward optimization (any reward at test time), from ONE policy ๐Ÿค–Natural recovery & transition
7
74
23
272
84,337
Teaching robots sub-millimeter precision through spatially-grounded touch We introduce SaTA which anchors tactile to hand kinematics, enabling learning-based policies to match model-based precision. arxiv.org/abs/2510.14647
5
37
4
189
45,998
Ever want to enjoy all the privileged information in sim while seamlessly transferring to the real world? How can we correct policy mistakes after deployment? ๐Ÿ‘‰Introducing GSWorld, a real2sim2real photo-realistic simulator with interaction physics with fully open-sourced code.
8
67
19
301
87,228
Rollouts in the real world are slow and expensive. What if we could rollout trajectories entirely inside a world model (WM)? Introducing ๐Ÿš€Ctrl-World๐Ÿš€, a generative manipulation WM that can interact with advanced VLA policy in imagination. ๐Ÿงต1/6
5
40
6
210
96,469