Step</th>	Loss</th>	Grad Norm</th>	LR</th></tr></thead>
5k</td>	0.011</td>	0.18</td>	~8.5e-5</td></tr>
10k</td>	0.010</td>	0.19</td>	7.6e-5</td></tr>
15k</td>	0.007</td>	0.13</td>	~4.8e-5</td></tr>
20k</td>	0.005</td>	0.11</td>	2.7e-5</td></tr> </tbody></table> Results #</a> </span> </h3> The v3 dual-camera model typically achieves 60-80% success across 5-episode eval runs. The arm approaches the cube, aligns its wrist, grasps, lifts, carries to the bowl, and releases. There's some shaking during execution, but this is a common behavior with IL models. I feel the shaking and slightly lower success rate might also be due to the time I performed inference. I recorded the dataset in late afternoon toward 5pm in late February, but during the time of evaluation it was around 12pm, which can slightly affect the lighting condition even though I have the dominant lighting (desktop lamp).</p> Failure modes are mostly spatial: occasional off-center grasps where the cube slips, or slight misjudgment of the bowl position on placement. The policy clearly understands the task structure and it's not randomly flailing.</p> Here's a demo video showing a successful run. One of the episodes had a funny moment where the robot fumbled the cube and then recovered, which was very rare.</p> Yet another VLA inference vid, but I fine-tuned SmolVLA on 75 demos for 20k steps on my RTX 3090. ~10.4 hours of training, usually 60-80% success rate. Way less painful than HIL-SERL from scratch, but now the performance is capped by human demos. Can you fine-tune VLAs with RL? pic.twitter.com/qBkHjONibx</a></p>— Gota (@gtgando) February 22, 2026</a></blockquote>

Gota Gando

Escaping the Tokyo Summer: Melbourne and Bali

2026-09-02T00:00:00+00:00

I tend to travel in August since Tokyo is very hot and humid during the summer, and I'm not productive if I stay here anyway. I usually procrastinate the planning until the last moment and feel annoyed about starting the travel, but for someone like me who works on a computer all day long, it always feels worth it after I've done it.</p>

Last year, I first visited SF, where I learned that robotics is getting hot, and then vacationed in Vancouver and Seattle. I considered traveling to the east coast this summer, but the US is too expensive for yen earners these days. So I decided to travel south this year.</p>

Melbourne #</a> </span> </h3>
First, I went to Melbourne for a week to escape the heat in the Australian winter. I mostly stayed in the Central Business District (CBD), and my first impression was that it's pretty clean and safe despite the size of the city. Also, there were so many Asian tourists that I kept hearing Japanese around the CBD. Makes sense given how close Australia is to Asia, and hey, I was one of them.</p>
Remote working #</a> </span> </h4>
Another reason I chose Melbourne was that I thought it could be a good place for short-term remote working. I was able to work from my Airbnb since the room was pretty spacious and quiet, but I actually felt it's not so good for nomading as 1. it's relatively expensive to stay and 2. cafe seating tends to be compact and not suitable for laptops. So you'd end up going to libraries, but they're a bit crowded as everyone else tries to do the same. However, I did find two cafes where I was able to work a bit:</p>

Luna Coffee</a>: It's a spacious hotel lounge with a small coffee shop. It's not crowded at all and nobody cares even if you stay here for a few hours.</li>
Krimper Cafe</a>: There's enough seating that the staff probably don't mind a laptop user for a while (probably). I went there in the morning when it's not crowded and was able to work for like 1-2 hours.</li> </ol>
For the library, I recommend the City Library since you can probably find a spot there, instead of the overcrowded State Library.</p>
Food #</a> </span> </h4>
As for restaurants and stuff, you can just research with Claude, Grok, etc., plus there's a shit ton of travel content on the web obviously, but I'd recommend these two restaurants I really liked:</p>

Vespa Rossa Degraves</a>: Beef Ragu Risotto</li>
Claypots Barbarossa</a>: Salmon Belly</li> </ul>
Architecture and NGV #</a> </span> </h4>
Also I really liked the buildings in the CBD. Lots of towers with reflective glass, and my favorite was UNO Melbourne, whose facade shifts between purple and copper depending on the angle. At NGV I also found a funny oil painting from 1652 (A Mediterranean port scene by Jan Baptist Weenix) where the lady with the parasol looks straight up like an anime character.</p>
</div>
</div>
Tech scene #</a> </span> </h4>
Overall the climate is pretty chill and people are friendly and great, but I didn't feel much impact in terms of networking. Nothing on the meetup scene looked that compelling to me personally, but I attended two anyway. I felt like I met more people who have actually built something and are actively working on it even in Vancouver, people you'd genuinely feel like you wanna connect with. Claude mentioned Melbourne kinda has a culture of indie hackers, but I didn't see anything particularly strong on that front. I stayed in Van for two weeks compared to one week in Mel, and the timing obviously matters, so I'm pretty biased here. For tech events I believe North America clearly wins, but for tourism and chill remote working, Melbourne was definitely worth it.</p>
Bali #</a> </span> </h3>
After Australia, I visited Bali for another week, basing myself in Canggu, to 1. do something summery since I'm getting hella old and 2. check out the digital nomad scene. I had such low expectations of Bali because people online kept trashing it, like how overcrowded, dirty, and expensive it has gotten compared to the past. But I loved Bali!</p>
There are some spots that are dirty, and there were indeed lots of tourists in the peak dry season. But I enjoyed it. Beaches are absolutely beautiful and worth visiting just to view them, even if you don't swim or surf.</p>
</div>
I couldn't explore everything, but there are lots of temples and world heritage sites you wanna visit, like Tanah Lot, the Rice Terraces and Uluwatu Temple. I was planning to just work on coding projects at nomad cafes, but ended up doing full tourist stuff.</p>
</div>
I went to Uluwatu just for the temple this time. Monkeys were literally stealing iPhones and glasses from tourists right in front of me, and the Kecak fire dance was really good, honestly better than the cliff views. For Ubud I did the classic route: the Tegallalang Rice Terraces → Tirta Empul → Serayu pottery → luwak coffee tasting → the Campuhan Ridge Walk. And Tanah Lot was definitely the best spot to view waves, with the temple sitting on a rock just offshore and the surf crashing around it. However touristy it looks, it's worth going at least once.</p>
Also, things are very cheap even for those who carry yen. For example, even lunch/dinner at a high-grade restaurant only costs like less than 3k yen, which is like half of what it'd be in Tokyo. Basically I felt like an American visiting Japan in terms of cost of living.</p>
Getting around #</a> </span> </h4>
For transportation, if you have a license for scooters it's definitely better to rent a scooter, though the traffic is pretty chaotic and you'd need decent skill to maneuver without hitting something. I technically have a Japanese driver's license that comes with a small scooter license up to 75cc, but Bali scooters are usually 125cc. This means it's technically illegal for me to drive them and travel insurance won't cover it, so I opted for just using the Gojek / Grab apps. I heard some people don't care and drive them anyway without the license, which I'm not really willing to risk. I ended up getting along with one of the car drivers I met on Gojek, and used his private service throughout the rest of my stay. Gojek takes a 20% margin, so if you pay the driver in cash plus some tips, it'd be a great deal for the driver. For reference, Canggu to Uluwatu Temple was around 400k IDR one way, and Canggu to the Tegallalang rice terraces was about 332.5k on Gojek, so Ubud day trips end up a bit more. After exchanging WhatsApp, we just used the Gojek app to check the fare for each ride and I paid him that in cash plus some tips. I'd recommend doing the same, for example if you stay around Canggu and plan to do long day trips to places like Ubud and Uluwatu.</p>
The nomad scene #</a> </span> </h4>
As for the nomad scene, I did enjoy the co-working cafes designed for laptops. It wasn't overcrowded and I could always find a spot. Working on your projects while sipping tropical drinks has its own charm.</p>
</div>
But I felt like it's kinda hard to connect with other fellow nomads / remote workers, because everyone kinda works on their laptops quietly. People were polite though, like you could ask people to watch your stuff for example. You probably wanna be proactive if you want to socialize with nomads. I'd attend events hosted by coworking spaces like Tribal. I didn't, since I wanted to focus on traveling around Bali. There was also a startup event on Eventbrite I wanted to attend, but it was sold out when I tried to buy a ticket! For the record, here are the ones I tried out:</p>

IZE</a>: very quiet and suitable for actual working</li>
Tribal</a>: the vibe is great but it's pretty busy. I suppose you'd need to come in the early morning to secure a spot</li>
NOAH Cafe</a>: there's a little enclosed space with AC for working, and it's pretty good</li>
ZIN Cafe</a>: arguably the best one in Canggu, plenty of seating, you can order great foods and drinks while working and they'll bring them to you</li> </ul>
People #</a> </span> </h4>
My impression from my limited stay was that tourists (also nomads?) are a bit indifferent to each other and just do their own thing alone or with their group, though most of the ones I encountered were polite. On the other hand, I had a much easier time connecting with local people, since people tend to be warm and friendly. Also, some people kinda speak Japanese and seem to like Japan too. Classic, but people seem to like anime and manga apparently (soft power is real). I think Balinese are friendly and they tend to speak good English, sometimes even Japanese. Communicating in English, Japanese, and my occasional bad Indonesian was pretty fun!</p>
Next time #</a> </span> </h3>
Overall, I definitely enjoyed Bali and would like to stay again, preferably a bit longer, in the near future. For anyone wondering if Bali is worth it, I'd personally recommend it. It is crowded with tourists, but it's famous for a good reason. I only briefly visited Uluwatu/Ubud this time, so next time I'm hoping to explore a bit north like Mt. Batur/Kintamani and toward the east like Nusa Penida, Gili, and Lombok.</p>

Fine-tuning SmolVLA for Cube Pick-and-Place on SO-101

2026-03-01T00:00:00+00:00

Context #</a> </span> </h3>
For the past few months I've been exploring different algorithms for my SO-101 robot arm to do a pick-and-place task. Last time I trained SAC with HIL-SERL (Human-in-the-Loop Sample Efficient RL) on the real robot with a reward classifier and human interventions via the leader arm. After about 750 episodes of training, HIL-SERL achieved 80% success on reach-and-grasp. That's good, but each episode required me to physically be there watching the robot, ready to intervene, and it's very labor-intensive.</p>
So this time I tried out VLAs (Vision-Language-Action models) to see how quickly I could train a working model. I decided to go with SmolVLA from HuggingFace, as it's a compact VLA built on SmolVLM that you can fine-tune on consumer hardware (in my case, RTX 3090). The idea is simple: collect teleoperation demos, fine-tune the pretrained model, and deploy. No reward engineering, no sim-to-real gap, no sitting next to the robot for hours pressing buttons. Here's what I tried out.</p>
The Setup #</a> </span> </h3>
Hardware: SO-101 follower + leader arm (Feetech STS3215 servos), Intel RealSense D405 as wrist camera, Logitech C920 as overhead camera, and my RTX 3090 for training. Software: HuggingFace's LeRobot</a> framework with a custom fork that adds placo-based IK for automated resets between episodes.</p>
The task: pick up a small cube from a randomized position and place it in a bowl. 10-second episodes, 30fps, recorded via teleoperation with the leader arm.</p>
SmolVLA vs ACT #</a> </span> </h3>
SmolVLA is a 500M+ parameter model, but you only fine-tune ~50M (the action expert + projections). The vision encoder (SigLIP) and language model (SmolLM2) stay frozen. This means the model already has a strong visual prior from pretraining. It knows what objects look like, it just needs to learn what to do with them.</p>
I also trained an ACT baseline for comparison. ACT is a solid architecture but had practical issues on my setup: no built-in image resize meant 640×480 frames produced 602 encoder tokens per forward pass, OOM'd at batch_size=64 on 24GB VRAM, and required double the training steps to match SmolVLA's sample count. At 1.28M matched samples, SmolVLA's L1 loss converged to 0.006 while ACT was at 0.046 (though the losses aren't directly comparable since ACT includes KL divergence).</p>
Three Rounds of Data Collection #</a> </span> </h3>
I had to record three times due to hardware breakage and inconsistent grasp motions:</p>
v1: 30cm workspace, 50 episodes #</a> </span> </h4>
First attempt. Cube placed randomly across a ~30cm area, two cameras (wrist + overhead). Training converged (loss dropped from 0.162 to 0.005 in 20k steps), but at inference the arm moved toward the cube and then grasped at air. 50 demos across 30cm just wasn't enough density. The policy learned the general motion but couldn't pin down precise grasp locations.</p>
v1 redux: 10cm workspace, 75 episodes #</a> </span> </h4>
Narrowed the workspace to ~10cm and collected 75 clean episodes. Also swapped the wrist camera from InnoMaker RGB (which kept dying mid-session) to a RealSense D405. This version hit 80% success on the first eval run. The tighter workspace gave the policy enough demonstration density to learn reliable grasping. I should have recorded the demo video and posted it at this point, but I realized that the flexible finger of the gripper was half-broken. I replaced it with a new one, re-calibrated, and obviously the performance significantly dropped since the model was trained with the old calibration. I decided to redo the entire process again.</p>
v2: The nudge trick problem #</a> </span> </h4>
Collected 81 more episodes with a recalibrated arm. During recording I developed a habit of nudging the cube with the gripper's static finger to rotate it into a better angle before grasping. Seemed clever at the time.</p>
The overhead camera also died partway through recording (likely USB power brownout from sharing the bus with the RealSense and servo controllers), so I bought a new one. Training metrics looked identical to v1: same final loss, same convergence. But eval was pretty bad at 20-80% success rate with wild variance across runs. The policy learned a conditional behavior: sometimes nudge, sometimes grasp directly. It couldn't consistently decide which to do.</p>
This was because I got used to the grasp task and started doing the nudging motion: nudge the cube to correct the rotation without having to rotate the wrist roll joint, then grasp. I also included another way to grasp, which is rotating the wrist roll to align the gripper angle and then grasping. This mix unfortunately turned out to be pretty bad for model training.</p>
Lesson learned: consistency in demonstrations matters more than quantity. One clean strategy beats a mix of tricks.</p>
v3: Clean demos with wrist roll alignment #</a> </span> </h4>
Re-recorded everything with a strict protocol. No nudging. Lower the arm above the cube, rotate the wrist roll to match the cube's angle, descend, grasp. Same strategy every time. Also replaced the dead C920 and moved the RealSense to a proper USB 3.0 port (it had been silently running at 480Mbps on a USB 2.0 port, and I found this with `lsusb -t</code>).</p>`
`75 clean episodes after removing 5 bad ones.</p>`
`Training`#</a> </span> </h3> All SmolVLA runs used the same hyperparams: batch_size=64, 20k steps, cosine decay from 1e-4, checkpoints every 5k steps. Images get resize_with_pad</code> to 512×512. Each run took about 10 hours on the 3090.</p> The v3 dual-camera run (wrist + overhead) converged to loss 0.005 at 20k steps. Loss was still dropping, so there's room to push further. Gradient norm went from 0.18 → 0.11, stable convergence throughout.</p> Step</th> Loss</th> Grad Norm</th> LR</th></tr></thead> 5k</td> 0.011</td> 0.18</td> ~8.5e-5</td></tr> 10k</td> 0.010</td> 0.19</td> 7.6e-5</td></tr> 15k</td> 0.007</td> 0.13</td> ~4.8e-5</td></tr> 20k</td> 0.005</td> 0.11</td> 2.7e-5</td></tr> </tbody></table> Results #</a> </span> </h3> The v3 dual-camera model typically achieves 60-80% success across 5-episode eval runs. The arm approaches the cube, aligns its wrist, grasps, lifts, carries to the bowl, and releases. There's some shaking during execution, but this is a common behavior with IL models. I feel the shaking and slightly lower success rate might also be due to the time I performed inference. I recorded the dataset in late afternoon toward 5pm in late February, but during the time of evaluation it was around 12pm, which can slightly affect the lighting condition even though I have the dominant lighting (desktop lamp).</p> Failure modes are mostly spatial: occasional off-center grasps where the cube slips, or slight misjudgment of the bowl position on placement. The policy clearly understands the task structure and it's not randomly flailing.</p> Here's a demo video showing a successful run. One of the episodes had a funny moment where the robot fumbled the cube and then recovered, which was very rare.</p> Yet another VLA inference vid, but I fine-tuned SmolVLA on 75 demos for 20k steps on my RTX 3090. ~10.4 hours of training, usually 60-80% success rate. Way less painful than HIL-SERL from scratch, but now the performance is capped by human demos. Can you fine-tune VLAs with RL? pic.twitter.com/qBkHjONibx</a></p>— Gota (@gtgando) February 22, 2026</a></blockquote>