Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
EgoBridge: Domain Adaptation for Generalizable Imitation from Egocentric Human Data
Authors: Ryan Punamiya, Dhruv Patel, Patcharapong Aphiwetsa, Pranav Kuppili, Lawrence Zhu, Simar Kareer, Judy Hoffman, Danfei Xu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Ego Bridge on both a reproducible simulation benchmark task and three challenging real-world manipulation tasks. Our results show that Ego Bridge consistently improves policy success rates compared to human-augmented cross-embodiment baselines, for up to 44% absolute success rate improvement, and effectively transfers behaviors from diverse human demonstrations to robotic execution in tasks requiring spatial, visual, and task generalization. |
| Researcher Affiliation | Academia | Ryan Punamiya1 Dhruv Patel1 Patcharapong Aphiwetsa1 Pranav Kuppili1 Lawrence Y. Zhu1 Simar Kareer 1* Judy Hoffman1* Danfei Xu1* 1Georgia Institute of Technology *Equal advising EMAIL |
| Pseudocode | Yes | A Algorithm Pseudocode Overview. Algorithm 1 describes the joint policy co-training of the human and robot data with the joint OT loss. ... Algorithm 1 Ego Bridge Co-Training for Human Robot Imitation |
| Open Source Code | No | Answer: [No] Justification: No. But we plan to release after acceptance. Guidelines: The answer NA means that paper does not include experiments requiring code. Please see the Neur IPS code and data submission guidelines (https://nips.cc/ public/guides/Code Submission Policy) for more details. While we encourage the release of code and data, we understand that this might not be possible, so No is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). The instructions should contain the exact command and environment needed to run to reproduce the results. See the Neur IPS code and data submission guidelines (https: //nips.cc/public/guides/Code Submission Policy) for more details. The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. |
| Open Datasets | No | Answer: [No] Justification: No. But we plan to release after acceptance. |
| Dataset Splits | No | The paper provides evaluation setup details such as: "Evaluation uses 48 trials (2 rollouts for each of the 24 drawers)", "Performance is measured by success rate over 15 rollouts across 5 distinct target locations", and "We conduct 18 evaluations with diverse shirt initial placement and colors". For simulation: "We evaluate a total of 100 fixed seeds across all the models". However, it does not explicitly state the training/validation/test splits of the collected human or robot demonstrations. |
| Hardware Specification | Yes | We train the real world Ego Bridge model on a single L40s gpu for 100000 iterations on the Drawer task, 110000 iterations on the Laundry task, and 120000 iterations on the Scoop Coffee task, which takes about 24 hours. ... We train the simulation Ego Bridge model on a single A40 GPU for 130000 iterations, which corresponds to around 2 hours of training time. |
| Software Dependencies | No | The paper mentions software components and tools like "Res Net-18 encoder", "Adam W" (optimizer), "Geom Loss [37]", "Sinkhorn algorithm [29]", "Diffusion Policy benchmark suite [6]", and "Co Tracker [35]". However, it does not provide specific version numbers for programming languages (e.g., Python), deep learning frameworks (e.g., PyTorch, TensorFlow), or other relevant libraries used for implementation. |
| Experiment Setup | Yes | Table 3: Hyperparameters for Real-World Experiments Symbol Value H W C 480 640 3 dproj 256 Dstem 8 dattn 64 L 16 dq(Robot) 7 or 14 (joint positions) dq(Human) 6 or 12 (end effector pose as xyz + euler) Dtrunk 8 Ntrunk 16 d 256 M 8 Dhead 8 Nhead 8 dhead 64 k 100 da 7 or 14 (xyz + euler angles + gripper position) LBC (Robot) Smooth L1(xyz) + Smooth L1(gripper) + 0.5 MSE(euler) LBC (Human) Smooth L1(xyz) L = LBC-cotrain + α LOT-joint α = 0.7 ... Table 4: Training Details for Real World Experiments Parameter Value Optimizer Adam W Learning Rate 5 10 5 Weight Decay 0.0001 Scheduler Linear Batch Size 32 Data Augmentations Color Jitter + Image Net Normalization (Res Net) OT Loss Geom Loss [37] Blur 0.05 Distance Sinkhorn ... Table 5: Training Details for Simulation Experiments (Push T) Parameter Value Optimizer Adam W Learning Rate 1 10 4 Weight Decay 1 10 6 Scheduler Cosine Warmup Steps 500 Iterations 130,000 Batch Size 32 Exponential Moving Average (EMA) Power = 0.75 Data Augmentations Image Net Normalization (Res Net) LBC (Triangle & Circle) MSE(xy) L = LBC-cotrain + α LOT-joint α = 0.2 OT Loss Geom Loss [37] Blur 0.01 Distance Sinkhorn |