Skip to content

Commit 522f703

Browse files
claudeclaude
committed
docs: #157 — research conference opportunities and draft GSoC 2027 application
Adds two documents to docs/academic/: - conference-opportunities-2026.md: researched status of ICRA 2026 (missed — workshops already announced), CoRL 2026 (actionable — abstract deadline May 25), NeurIPS 2026 (main track missed; watch for workshop CFPs in September), Google DeepMind Accelerator (missed for current cohort), and GSoC 2027 (org application window ~Jan–Feb 2027). - gsoc-2027-application.md: draft org application with five concrete project ideas (Gazebo/ROS2 backend, VLM judge, Isaac Lab integration, live dashboard, WBC benchmark suite), eligibility checklist, mentor roster template, and submission checklist. https://claude.ai/code/session_01EPHNDzZLoHVZsx1PmxXciG
1 parent 3dfa99e commit 522f703

2 files changed

Lines changed: 366 additions & 0 deletions

File tree

‎docs/academic/conference-opportunities-2026.md‎

Lines changed: 174 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,174 @@
1+
# Academic Conference Opportunities — 2026
2+
3+
_Research compiled: April 2026. Revisit for ICRA 2027 in Q4 2026; revisit for NeurIPS 2026 workshops in September 2026._
4+
5+
---
6+
7+
## Summary
8+
9+
| Venue | Dates | Deadline | Action | Status |
10+
|-------|-------|----------|--------|--------|
11+
| ICRA 2026 | Jun 1–5, Vienna | **Passed** | Plan ICRA 2027 instead | ❌ Missed |
12+
| CoRL 2026 | Nov 9–12, Austin TX | Abstract May 25 / Paper May 28 | System paper possible | ⚡ Actionable |
13+
| NeurIPS 2026 (main) | Dec 6–12 | May 6 AOE | **Passed** | ❌ Missed |
14+
| NeurIPS 2026 (workshops) | Dec 6–12 | ~September (TBA) | Monitor, submit demo/position paper | ⏳ Watch |
15+
| Google DeepMind Accelerator | Cohort Jun 2026 | Closed Mar 25 | Monitor for next cohort | ❌ Missed, next TBA |
16+
| GSoC 2027 | Summer 2027 | Org apps ~Jan–Feb 2027 | Prepare materials Q4 2026 | ⏳ Plan |
17+
18+
---
19+
20+
## ICRA 2026 — Missed, Plan for 2027
21+
22+
**Dates:** June 1–5, 2026 (workshops: June 1 and June 5), VIECON, Vienna, Austria
23+
**Status:** Workshop notifications already sent. All workshop submission deadlines have passed.
24+
25+
### Relevant workshops (for reference and ICRA 2027 planning)
26+
27+
The following workshops at ICRA 2026 are closely aligned with roboharness:
28+
29+
| Workshop | Day | Relevance |
30+
|----------|-----|-----------|
31+
| Synthetic Data for Robot Learning | Mon Jun 1 | Simulation-based data, visual validation |
32+
| Generative Digital Twins for Sim2Real and Real2Sim Transfer | Mon Jun 1 | Sim-to-real validation gap — exactly what roboharness addresses |
33+
| Reinforcement Learning in the Era of Imitation Learning | Mon Jun 1 | Evaluation methodology for RL policies |
34+
| From Data to Decisions: VLA Pipelines for Real Robots | Fri Jun 5 | VLA evaluation, tool-chain gap |
35+
| Beyond Teleoperation: Learning from Diverse Human and Simulation Data | Fri Jun 5 | Sim-based learning evaluation |
36+
37+
**ICRA 2027 action plan (Q4 2026):**
38+
39+
- Identify call for workshop proposals (typically released ~October, deadline ~December)
40+
- Submit a **workshop proposal** for "Visual Evaluation Frameworks for Robot Learning Agents"
41+
- Angle: AI coding agents need visual feedback loops; existing CI for robotics is blind
42+
- Roboharness positions as a reference implementation of the harness engineering approach
43+
- Alternatively, submit a **demo paper** to an aligned workshop (Sim2Real or RL evaluation)
44+
45+
---
46+
47+
## CoRL 2026 — Actionable (Abstract Due May 25)
48+
49+
**Dates:** November 9–12, 2026, Austin, Texas, USA
50+
**Abstract deadline:** May 25, 2026
51+
**Full paper deadline:** May 28, 2026
52+
**Submission:** https://openreview.net/group?id=robot-learning.org/CoRL/2026/Conference
53+
54+
### Fit assessment
55+
56+
CoRL accepts papers on robot learning with real-robot experiments or strong sim-to-real evidence. A roboharness **system paper** is plausible if:
57+
58+
1. Experimental validation with a real robot (or convincing sim-to-real transfer evidence) is available
59+
2. The paper demonstrates improvement in an agent's iteration speed or debugging quality
60+
3. User study or quantitative comparison with baseline (no visual feedback) exists
61+
62+
**Current gap:** Roboharness lacks the real-robot experiments CoRL requires. Purely sim-based results are unlikely to clear review bar without sim-to-real justification.
63+
64+
### Recommended paper angle (if experiments are available)
65+
66+
**Title:** "Harness Engineering for Robot Learning: Visual Checkpoint Testing Accelerates AI Agent Iteration"
67+
68+
**Key claims:**
69+
- AI coding agents editing robot control policies benefit from visual feedback at checkpoint intervals
70+
- Checkpoint screenshots + structured JSON reduce debugging loop time vs. log-only iteration
71+
- Integration is three lines of code (drop-in `RobotHarnessWrapper`)
72+
- Demonstrated on: MuJoCo Grasp, G1 WBC Reach, G1 LeRobot locomotion
73+
74+
**Decision checklist (before submitting):**
75+
- [ ] Can we generate quantitative data comparing agent iteration speed with/without roboharness?
76+
- [ ] Do we have real-robot or strong sim-to-real experiment?
77+
- [ ] Is the paper 8 pages achievable given current result set?
78+
79+
**If yes to all → submit.** Deadline May 25 for abstract.
80+
81+
---
82+
83+
## NeurIPS 2026 — Main Track Missed; Watch for Workshops
84+
85+
**Dates:** December 6–12, 2026
86+
**Main track full paper deadline:** May 6, 2026 AOE — **passed**
87+
**Workshop CFPs:** Typically announced September 2026 (check neurips.cc/Conferences/2026)
88+
89+
### Workshop strategy
90+
91+
NeurIPS workshops are a better fit for roboharness than the main track because:
92+
- Lower bar for tool/framework papers without extensive real-robot experiments
93+
- Position papers and demos are common
94+
- Robotics × ML × evaluation methodology is a natural workshop fit
95+
96+
**Candidate workshop topics to watch for in September 2026:**
97+
- "Robot Learning" or "Embodied AI" workshop
98+
- "Evaluation and Benchmarking in Robotics" (recurring topic)
99+
- "Machine Learning for Robotics" or "Foundation Models for Robotics"
100+
- "AI for Scientific Discovery" (sim-based agent testing fits here)
101+
102+
**When workshops are announced:**
103+
1. Identify 2–3 workshops with deadlines in October–November 2026
104+
2. Submit a **4-page demo/position paper** describing roboharness + results from showcase demos
105+
3. Angle: "Harness engineering as the missing CI layer for robot learning"
106+
107+
---
108+
109+
## Google DeepMind Accelerator: Robotics
110+
111+
**Status:** First cohort applications closed March 25, 2026. Cohort kicks off June 2026 in London.
112+
113+
**Program details:**
114+
- 10–15 early-stage robotics startups (European HQ, VC-backed, 5+ technical team members)
115+
- Equity-free, 12–15 weeks
116+
- Up to $350K Google Cloud credits via Google for Startups
117+
- Target sectors: logistics, manufacturing, health/life sciences, HRI, education, navigation
118+
119+
**Current eligibility gap:**
120+
- roboharness is an open-source project, not a VC-backed startup — does not meet current eligibility criteria
121+
- The program targets commercial startups, not OSS/academic tools
122+
123+
**Action:** Monitor for future cohorts or a research track. Check https://deepmind.google/models/gemini-robotics/accelerator/ in Q3/Q4 2026 for next cohort announcement.
124+
125+
---
126+
127+
## GSoC 2027 — Prepare Materials Q4 2026
128+
129+
**Program:** Google Summer of Code — open-source organization sponsoring student contributors
130+
**Org application window:** Typically January 22 – February 6 (based on 2026 timeline; 2027 dates not yet published)
131+
**Accepted orgs announced:** ~February 21 (based on 2026)
132+
133+
### Eligibility checklist
134+
135+
- [x] Open-source project (MIT license)
136+
- [x] Active GitHub repository with meaningful commit history
137+
- [x] Clear contributing guide (CONTRIBUTING.md exists)
138+
- [ ] Prior GSoC participation (not required for first-time orgs, but helpful)
139+
- [ ] Minimum 2 mentors willing to commit time (need to identify)
140+
141+
### What to prepare (Q4 2026, target November–December)
142+
143+
See `docs/academic/gsoc-2027-application.md` for the full draft application.
144+
145+
**Key sections needed for org application:**
146+
1. Organization name and description (2–3 sentences)
147+
2. Why Google should accept roboharness as a GSoC org
148+
3. Project ideas list (3–5 concrete ideas with scope, skills required, difficulty)
149+
4. Mentor roster (name, GitHub handle, availability)
150+
5. Communication channels (GitHub Discussions, Discord, etc.)
151+
152+
**Project idea candidates:**
153+
- **Gazebo/ROS2 backend** — implement `SimulatorBackend` for ROS-based Gazebo simulation
154+
- **VLM evaluation judge** — integrate open-source 4B VLM for autonomous checkpoint scoring
155+
- **Isaac Lab checkpoint integration** — full round-trip with NVIDIA Isaac Lab, visual reports
156+
- **Web dashboard** — real-time Rerun/Meshcat-based live monitoring dashboard
157+
- **Pinocchio/MuJoCo benchmark suite** — standardized evaluation harness for WBC controllers
158+
159+
**Timeline:**
160+
- Q3 2026 (Jul–Sep): Build community presence (ROS Discourse, HuggingFace, Discord)
161+
- Q4 2026 (Oct–Dec): Draft project ideas, identify mentors, prepare application
162+
- Jan 2027: Submit org application when window opens
163+
- Feb 2027: If accepted, publish project ideas and start engaging potential contributors
164+
165+
---
166+
167+
## References
168+
169+
- ICRA 2026: https://2026.ieee-icra.org/
170+
- CoRL 2026: https://www.corl.org/
171+
- NeurIPS 2026: https://neurips.cc/Conferences/2026
172+
- Google DeepMind Accelerator: https://deepmind.google/models/gemini-robotics/accelerator/
173+
- GSoC timeline (2026 reference): https://developers.google.com/open-source/gsoc/timeline
174+
- Issue: miaodx/roboharness#157

‎docs/academic/gsoc-2027-application.md‎

Lines changed: 192 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,192 @@
1+
# GSoC 2027 — Organization Application Draft
2+
3+
_Status: Draft — finalize and submit when org application window opens (~January 2027)_
4+
_Reference timeline: GSoC 2026 org window was Jan 22 – Feb 6; accepted orgs announced ~Feb 21_
5+
6+
---
7+
8+
## Organization Profile
9+
10+
**Organization name:** Roboharness
11+
12+
**Short description (≤ 180 chars):**
13+
Visual testing harness for AI coding agents in robot simulation. Helps Claude Code and Codex see what robots are doing and iterate autonomously.
14+
15+
**Long description:**
16+
17+
Roboharness is the first harness engineering tool built specifically for robotics. When AI coding agents (Claude Code, Codex) write or debug robot control code, they are blind — they can read logs and numbers but cannot see what the robot is actually doing. Roboharness solves this by capturing multi-view screenshots and structured state JSON at user-defined checkpoints, giving agents the visual context they need to iterate autonomously.
18+
19+
The project supports MuJoCo, Gymnasium (zero-change `RobotHarnessWrapper`), Isaac Lab, LeRobot G1 locomotion, and SONIC motion tracking. It produces auto-generated HTML visual reports that CI can publish, and includes an MCP server so agents can query checkpoint results directly. The project is MIT-licensed, Python 3.10+, and follows a SimulatorBackend protocol for easy extension to new simulators.
20+
21+
Roboharness is where robotics meets harness engineering — a category that until now existed only for software, not physical systems.
22+
23+
**Website:** https://github.com/MiaoDX/roboharness
24+
**License:** MIT
25+
**Primary language:** Python
26+
**Tags:** robotics, simulation, testing, AI agents, MuJoCo, Gymnasium, evaluation
27+
28+
---
29+
30+
## Why Should Google Accept Roboharness?
31+
32+
1. **Category-defining project.** Roboharness occupies a unique niche: visual regression testing for robot simulation. No other open-source tool targets this exact intersection of AI coding agents + robot simulation + visual feedback.
33+
34+
2. **Growing relevance.** AI coding agents (Claude Code, Codex) are now writing real robot code. The tooling to support them — especially visual testing and evaluation — doesn't exist yet. Roboharness fills that gap.
35+
36+
3. **Student-friendly architecture.** The `SimulatorBackend` protocol (structural typing, no inheritance) makes adding a new simulator backend an ideal GSoC project: well-scoped, testable, standalone, and directly useful to the community.
37+
38+
4. **Active development.** Regular commits, CI with real robot simulation tests, live HTML reports auto-generated on every push.
39+
40+
5. **Mentor capacity.** Two core mentors available (see Mentors section), with domain expertise in robotics simulation, Python, and AI agent tooling.
41+
42+
---
43+
44+
## Project Ideas
45+
46+
### 1. Gazebo / ROS 2 Backend (Medium, 350 hours)
47+
48+
**Goal:** Implement a `SimulatorBackend` for ROS 2 / Gazebo, letting roboharness capture checkpoint screenshots from a running Gazebo simulation via ROS topics and TF.
49+
50+
**Why it matters:** Gazebo is the most widely used simulator in the ROS ecosystem. A ROS 2 backend would make roboharness accessible to thousands of ROS developers.
51+
52+
**Skills:** Python, ROS 2 (rclpy), Gazebo, basic image processing
53+
**Difficulty:** Medium
54+
**Deliverables:**
55+
- `roboharness/backends/ros2_gazebo.py` — implements `SimulatorBackend` protocol
56+
- Subscribes to `/camera/image_raw` (or similar) ROS topic for screenshot capture
57+
- Gets robot state via `tf2_ros`
58+
- Integration tests using `roboharness[dev]` (mock ROS publisher)
59+
- Example: `examples/ros2_gazebo_example.py`
60+
- Documentation
61+
62+
**Mentor:** TBD
63+
64+
---
65+
66+
### 2. VLM Evaluation Judge (Hard, 350 hours)
67+
68+
**Goal:** Add an optional `VLMJudge` evaluator that scores checkpoint screenshots using an open-source vision-language model (e.g., a 4B-parameter robot-specific VLM).
69+
70+
**Why it matters:** Currently, pass/fail evaluation requires hand-written constraint functions. A VLM judge would enable natural-language task specifications like "is the robot holding the cube?" without writing code.
71+
72+
**Skills:** Python, HuggingFace Transformers, vision-language models, evaluation methodology
73+
**Difficulty:** Hard
74+
**Deliverables:**
75+
- `roboharness/evaluate/vlm_judge.py` — pluggable VLM backend with a common interface
76+
- Support for at least two backends: local model (HuggingFace) + API model (Claude via anthropic SDK)
77+
- Prompt templates for common robot tasks (grasp success, reach target, avoid collision)
78+
- Benchmarking script comparing VLM scores vs. ground-truth constraint evaluators
79+
- Tests using mocked VLM responses
80+
- Documentation
81+
82+
**Mentor:** TBD
83+
84+
---
85+
86+
### 3. Isaac Lab Checkpoint Integration (Medium, 175 hours)
87+
88+
**Goal:** Complete the Isaac Lab integration with full round-trip checkpoint capture and auto-generated HTML visual reports, matching the quality of the MuJoCo backend.
89+
90+
**Why it matters:** NVIDIA Isaac Lab is the dominant GPU-based robot learning platform. Full roboharness support would make it accessible to Isaac Lab users without changing their training/eval code.
91+
92+
**Skills:** Python, NVIDIA Isaac Lab, PyTorch (basic), understanding of `SimulatorBackend` protocol
93+
**Difficulty:** Medium
94+
**Deliverables:**
95+
- `roboharness/backends/isaac_lab.py` — full implementation (currently partial)
96+
- Round-trip: observation → checkpoint → screenshot → structured JSON → HTML report
97+
- `examples/isaac_lab_harness_example.py` using a standard Isaac Lab environment
98+
- CI smoke test (mocked; real GPU test in separate CI tier)
99+
- Documentation
100+
101+
**Mentor:** TBD
102+
103+
---
104+
105+
### 4. Web Dashboard for Live Monitoring (Medium, 350 hours)
106+
107+
**Goal:** Build a lightweight web dashboard that shows live checkpoint screenshots and state data during a running simulation, using Rerun or a simple WebSocket server.
108+
109+
**Why it matters:** Currently, roboharness produces static HTML reports after the fact. A live dashboard would let agents (and humans) monitor robot behavior in real time.
110+
111+
**Skills:** Python, WebSocket, HTML/CSS/JS (basic), Rerun SDK or similar
112+
**Difficulty:** Medium
113+
**Deliverables:**
114+
- `roboharness/server/` — lightweight async WebSocket server (Python)
115+
- Live screenshot streaming (JPEG/PNG over WebSocket)
116+
- Static HTML frontend with auto-refreshing image grid + state JSON panel
117+
- `harness serve` CLI command
118+
- Tests (mock WebSocket client)
119+
- Documentation
120+
121+
**Mentor:** TBD
122+
123+
---
124+
125+
### 5. Benchmark Suite for Whole-Body Controllers (Medium, 350 hours)
126+
127+
**Goal:** Create a standardized benchmark suite that uses roboharness checkpoints to evaluate WBC (whole-body control) policies, producing comparable metrics across different controllers.
128+
129+
**Why it matters:** There is no standard benchmark for WBC evaluation. roboharness checkpoint data (screenshots + joint state + task success) is the right primitive for building one.
130+
131+
**Skills:** Python, MuJoCo, robotics (basic kinematics), statistics
132+
**Difficulty:** Medium
133+
**Deliverables:**
134+
- `roboharness/benchmark/` — benchmark runner + metric collection
135+
- At least 3 task scenarios: reach, grasp, balance
136+
- Per-scenario pass/fail + summary report
137+
- Comparison across at least 2 controllers (e.g., Pinocchio+Pink WBC vs. scripted baseline)
138+
- HTML report with per-task visual breakdown
139+
- Documentation
140+
141+
**Mentor:** TBD
142+
143+
---
144+
145+
## Mentors
146+
147+
_Note: Finalize mentor list before submitting. Each mentor must have a GitHub account and agree to the GSoC mentoring commitment (~5 hours/week during coding period)._
148+
149+
| Name | GitHub | Expertise | Availability |
150+
|------|--------|-----------|--------------|
151+
| TBD | @TBD | Robotics simulation, Python | TBD |
152+
| TBD | @TBD | AI agents, ML tooling | TBD |
153+
154+
**Minimum required:** 2 mentors (1 primary + 1 backup per project)
155+
156+
---
157+
158+
## Communication
159+
160+
- **GitHub Issues/Discussions:** https://github.com/MiaoDX/roboharness/discussions (primary channel)
161+
- **Discord:** TBD (set up before org application)
162+
- **Mailing list:** TBD (Google Groups or similar)
163+
164+
---
165+
166+
## Application Checklist
167+
168+
Before submitting the org application:
169+
170+
- [ ] At least 2 confirmed mentors with GSoC accounts
171+
- [ ] Discord server set up with `#gsoc-2027` channel
172+
- [ ] GitHub Discussions enabled and active
173+
- [ ] Project ideas page published on GitHub Wiki or docs site
174+
- [ ] CONTRIBUTING.md updated with GSoC-specific onboarding section
175+
- [ ] At least 3 merged contributor PRs from non-core team (shows welcoming community)
176+
- [ ] README has "Contributing" and "Community" sections prominently visible
177+
178+
---
179+
180+
## Submission Steps (when window opens ~Jan 2027)
181+
182+
1. Go to https://summerofcode.withgoogle.com/
183+
2. Sign in with a Google account associated with a GitHub account that has org admin access to MiaoDX/roboharness (or the `roboharness` org)
184+
3. Click "Apply as a Mentoring Organization"
185+
4. Fill in the profile using the sections above
186+
5. Submit before the deadline (~February 6, 2027 based on 2026 timeline)
187+
6. Monitor for acceptance announcement (~February 21, 2027)
188+
189+
---
190+
191+
_See also: `docs/academic/conference-opportunities-2026.md` for conference strategy_
192+
_Issue: miaodx/roboharness#157_

0 commit comments

Comments
 (0)